Changelog · v1.5.0
All notable changes to Airbuild are documented here.
Latest
Claude Sonnet 5.5 takes over as the default model and Claude Haiku 5.5 as the fast one beside it. Mistral Large 4 replaces Mistral Medium 3.5 in the picker. Haiku 5.5 costs a tenth of what Haiku 4.5 did per token and reads far more at once. A project pinned to one of the previous models keeps running on it — nothing is switched under you.
Unpublishing used to do two opposite things at once — the logs and the custom domain went immediately, the database stayed forever. Now everything stays for 30 days: database, logs and custom domain. Only the public address is dropped. Publishing again inside the window cancels the deletion, and a reminder goes out 23, 16 and 2 days before. Deleting a publication outright removes its data with it, in one step, and the entry only disappears once the data is really gone.
Before a publish, the gate lists every endpoint reachable without a login. Until now there were two honest answers: secure them, or publish anyway. There is a third now — tick the ones that belong to the public and the AI marks them in your code as deliberately public. That marker is what the gate reads on the next run, so the decision lives in your repository instead of in a dialog nobody can find again, and the publish restarts by itself once the change is committed. "Publish anyway" stays for the cases where the finding is simply wrong.
While the AI repairs a failed publish, the findings from the failed attempt no longer stand on screen as though nothing had happened. And the data export is back where it belongs, in the publish dialog.
Every row in the model list carried four letters — A, L, M, H — which made the list wider and explained nothing. Reasoning effort is its own chip next to the model picker now: it shows the active model’s level as an icon and a word, and it is tinted as soon as that level differs from the model’s default. The panel behind it explains each level in one line. The setting is still kept per model and per person, and a model with no adjustable reasoning simply gets no chip.
The agent runs the tests again to check itself, so a single message often carried five or more runs — stacked as near-identical cards. From the second run on they now collapse into one line ("Checked 7 times — everything works · 1m 22s") that opens into the individual runs, and each run into the names of its checks. A failure stays visible until a later run of the same scope answers it.
A site published without a domain of your own used to get an address at xaio.app. New publishes get one at airbuild.space, and the links stored for sites that already existed moved along with them. Nothing breaks on the way: the old addresses keep serving every site they ever served, and a custom domain is untouched.
While backend services are being replaced, an amber note at the top of the workspace says which areas are being updated. It warns, it does not block — you can keep working, and the note can be dismissed for that window. If a request fails during the rollout it comes back by itself, and a tab that stayed open through the whole thing gets a short "update complete" with a reload button.
Rolling one project out to more than one target belongs to Suite Builder. Losing that plan used to leave every additional target serving — and billing. Now the plan decides: the extra targets go offline, their data is kept for 30 days, and the deployment list says they are suspended and why. Getting the plan back brings them back; letting the window run out deletes them along with their database.
The copy held a single request open for as long as it took, and anything past two minutes ran into the gateway’s timeout — whereupon the workspace deleted the new project’s entry while the copy was still running, leaving it orphaned. The copy runs in the background now: the dialog waits twenty seconds and opens the project if it is ready, otherwise it closes and the tile shows that copying is under way. The entry is only rolled back when the copy has definitely failed.
The list of connected integrations showed a bare green tick per row instead of the name of the integration — the view was reading a field the backend no longer sends. And a search through the code that finds nothing is a normal step in looking around, not a failure; its card had a red border as though something had broken. It is grey now, like every other routine step.
The database view now shows the number of tables next to the heading of the list — and while a filter is active, how many of them match.
Three views stayed dark whatever you had set: the log stream, the code viewer and the generated API documentation. In a light workspace the log panel was a black block with colours on it that could not be read, and SQL files came out unhighlighted. All three take the theme now.
A preview runs the code that was written for you, right next to the workspace that wrote it. Messages between the two are now addressed to the preview’s exact address instead of to whoever is listening, and every message is checked for where it came from before it is read. The session cookie was narrowed to the one host it belongs to. Nothing changes in how you work; it changes what a preview can reach.
The entry screen asked what you want to build and listed everything you had already built, both at once — whichever you came for, the other was in the way. Home is about starting now: the composer, the five projects you last touched, and a link to the rest. All projects keeps the full overview, and the mark in the sidebar leads back home from there. Two things along the way: premium design is simply on instead of a switch that only let you make your own project worse, and when you drag a file in, the card you are dropping onto is the target — not a full-screen layer above it that swallowed the drop.
An additional target — a staging environment next to production, say — now gets its own database the way the default target always had one, and keeps it for as long as the target exists: created with the target, kept across every publish, dropped only when the target is deleted. The views follow along. The database browser, the logs, the metrics, the publish status and the data export all show the target you have selected instead of whichever one came first.
You can hand a project a rules file while you are creating it — a Markdown or text file up to 100 KB saying how this project is to be built: coding standards, naming, things to stay away from. The rules apply to every AI chat in the project from the first turn on, not from the moment someone remembers to mention them.
If a card was refused, the invoice behind the top-up stayed open — and the guard against double charges then refused every further top-up from that account until a nightly job cleared it, up to two days later. It caught people at the worst possible moment: out of AI Tasks, trying to pay, and told only that an earlier top-up was still pending. A refused charge now discards its invoice, and the reason reaches you in the chat in Stripe’s own words, "insufficient funds" instead of a riddle — enough to pick a smaller pack or another card. A charge waiting on 3-D Secure is the exception: that one can still be completed, so its payment page is kept and offered as a link.
The prompt behind the spreadsheet suggestion asked for the columns to become fields, which describes a table and not an app. It names the rest now: formulas become the logic of the app instead of being stored as their results, validation lists become select fields, lookup sheets become tables of their own, and whether a cell is typed or calculated decides what stays editable. It also asks for the existing rows as real data, with their units and formats, and for the wording of the sheet itself — so the app speaks the language you already use.
Article 50(5) asks for the notice to be clear and legible no later than the first interaction. Ours had drifted to the foot of the start screen, below the project tiles and so below the fold on an ordinary display — a first prompt could be sent without it ever having been on screen. It travels with the composer again. The same paragraph asks for it to meet accessibility requirements, and the grey it was set in reached 1.79:1 against its background where 4.5:1 is the bar. It is the same muted tone the website uses now, at 4.98:1.
Claude Opus 5.5 takes the place of Opus 5, and GPT-6 Sol and GPT-6 Luna take the place of the 5.6 pair. Sol costs less and answers faster than the model it replaces, and its rating in the picker says so. Opus 5.5 keeps adaptive thinking and stays on European infrastructure. A project pinned to one of the previous models keeps running on it, priced as before — nothing is switched under you.
The card that appears when an Airbuild link is pasted into a chat or a feed used to be the homepage hero: an eyebrow, a two-line claim, a paragraph and a colour band. It is now the wordmark on a light grey ground and nothing else. In a conversation a claim reads as an ad, while a logo reads as a sender — and the link’s own title already says what the page is.
The contact and sign-up pages were delivered as an empty document and only appeared once JavaScript had run — fine in a browser, invisible to everything else that reads a page. Both now arrive complete. Alongside that, 33 page titles were too long to be shown in full in a search result and 58 descriptions were cut off mid-sentence; none are now. And the showcase gallery finally calls the product by its current name.
Every published app gets its own database schema, and the statement that creates it looked like an injection risk to the scanners that read your project — so projects rewrote it by hand, and a rewrite that drops a type cast Postgres needs never parses. No migration ran, and every publish of that app failed and rolled back. The scaffold now writes the form that is safe and also reads as safe, and projects already carrying the broken rewrite are repaired in place — committed in your own repository, so the change appears in the Git panel instead of out of nowhere.
A video was recognised only by the type the browser reported, and for some recordings the browser reports nothing — so the file counted as “other”, took the wrong size limit and showed a broken-media icon where the thumbnail belongs. The file name is read as well now, and a format the browser cannot decode gets a drawn film frame instead of a broken picture.
When a published site sent a visitor on to its canonical address — the same page with or without the trailing slash — everything after the question mark was dropped on the way. A campaign link, a shared filter, a link out of an e-mail: the visitor arrived without them. The redirect carries the query string now.
Fable 5, Opus 5 and Sonnet 5 can vary their own thinking from step to step. Their effort control now offers Adaptive and starts there, so a simple step stays cheap and a hard one gets the room it needs, instead of everything running at one fixed ceiling. Every other model keeps Low, Medium and High, with High as its default.
The Design Mode prompt bar now accepts files the way the chat input does: the paperclip, drag and drop, paste from the clipboard — a pasted screenshot is named with its timestamp — thumbnails with a remove button and an upload ring, and the upload into your Media Library starts the moment you attach.
The home page was rebuilt around what Airbuild is for — business software a team runs on, rather than a prototype to throw away. The product itself is on the first screen where a film used to play, the arguments are addressed to the person buying rather than to their developer, and the page is half the length it was. Text contrast was reworked over every background, and the German pages have share cards of their own instead of the site-wide fallback.
A frontend-only publish announced “checking endpoints”, sat in that step while the task started up, and afterwards recorded it as passed — for a check that has nothing to look at without a backend. The publish now names the step that really runs first.
Every model in the picker now carries its own Low / Medium / High control, so the levels are visible side by side and setting one neither switches model nor closes the menu. Low is the lowest level that still makes sense for coding work, Medium is what Airbuild sent until now, High is the model's maximum — and High is the new default, because the better answer is usually worth more than the cheaper one. Thinking is billed like the rest of a model's output, so Low is there for work that does not need it. The level belongs to you and to that model rather than to a project, so it follows you wherever you use it.
Opening an organisation now requires a Suite Builder plan of your own — which means your own seat is already paid for, so only the members you add are charged. The purchase card prices it that way and says why, the counter starts at two, and a new organisation starts with one seat instead of five places nobody had paid for. It also counts how many of your members already hold a licence of their own and says so before you buy: a seat is charged on top of a personal subscription, it does not replace it.
The bar could sit at 10 % “Analyzing code” for minutes, because on the new build path the steps behind it had no way to report back — and a bar that does not move is indistinguishable from a publish that has hung. Progress is relayed again, and within a step it creeps forward instead of standing still. The scanners count themselves off as they go.
Turns on GPT-5.6 ended in the middle of the work — a short wrap-up with steps of the plan still open, and no error card to explain it. On the interface the model was called through it cannot use tools and reason at the same time, and the tools always win, so it had been running without reasoning. The whole 5.6 family now runs on the interface where both work together.
A budget could be created and deleted and nothing in between, so correcting a limit meant deleting it and losing the usage counted against it. Name, monthly limit, alert threshold and whether it blocks or only warns can all be changed now. A second budget at the same scope is refused instead of created — two of them did not merely read confusingly, they made every check that consulted them fail.
The trash icon in the members table had nothing behind it, and it was offered even for the sole owner, who cannot be removed at all. It now does what it says, and it is no longer shown where the only possible outcome is an error. Role names are written out in both languages, and the role menus were rebuilt so they look and read the same in every browser.
A publish can stop twice for a decision: once for endpoints that are open without authentication, once for security findings that are advisory. Both dialogs could stand on screen at the same time, so Cancel in one and “Publish anyway” in the other were live together. Only one is asked now, and the second follows if it still applies. The question also reaches the person who started the publish — it was read from something written only after your session had loaded, so on a freshly opened tab it could stay silent. And cancelling the first question no longer marks a publish that is still running as cancelled.
Following the rename, the company behind Airbuild is registered as Airbuild FlexCo. The imprint, the terms, the data processing agreement and the privacy page carry the new name; the address and everything else about the entity are unchanged.
The website can carry a prompt you typed there through sign-up and into the workspace. The last step of that handover was missing, so the text was dropped on arrival and you landed in front of an empty composer. It now opens the create panel with your words in it, and the project title and type are worked out from them exactly as if you had typed them inside.
A screen recording could be uploaded, stored and shipped in your app — but not watched. Reading it fell through to the text reader, which answered “cannot read file” and spent a whole iteration saying nothing. The assistant now watches it as timestamped frames: a coarse pass to find the moment, then a closer look at the seconds around it. Frames that are identical collapse into one that says the picture did not move, which is a finding rather than a gap. Narration can be transcribed as well, because “watch — when I click here, the form resets” lives in the voice and no frame shows it. A silent recording is reported as silent, and a transcription that fails says so instead of passing for silence. Each look appears in the chat as its own card, naming the window and the frame rate it used.
Three ways an attachment could be lost on the way. Images were dropped entirely on Gemini — the model was handed the words “[Image attached]” in place of the picture and answered about things it had never seen, screenshots included. Documents were discarded on every model outside the Claude family, although their text had already been extracted and was sitting right there, so you attached a specification and got an answer written from nothing. And a picture’s label — the page number of a PDF, the timestamp of a video frame — was separated from the picture it belonged to, leaving the model to match them by counting. Pictures and documents arrive now, PDFs are read natively where the model can, and every picture keeps its label.
Whole areas of the workspace were hardcoded English — and in a few places hardcoded German, so one panel could mix both whatever your setting said. The chat and every tool card, the publish and deployment flow, the log viewer of a published app, the test cockpit, the Postgres manager, agent flow, pull requests and conflict resolution, the design editors, the marketplace and the import panels are all translated now. A permission or session error reads in your language too, instead of falling back to English.
A running turn blocked every other project in your account: a message in a second chat came back with “Another AI turn is already running for your account” until the first one had finished — half an hour, on a long turn. The lock guarded nothing that the database was not already guarding a layer below it, and it is gone. One turn at a time per chat, as before.
The rename has reached the addresses: the website and the workspace answer on airbuild.ai, and the previous ones redirect there. A session you had open kept working across the change.
A workspace preview was reachable by anyone who had its address — the running app, its assets, and the backend it talks to. A preview now checks that you are on the project before it opens: the workspace hands your browser a short-lived pass, and the preview host accepts nothing else. The preview tab, a preview opened in its own tab and the screenshots the assistant takes all work exactly as before. A link forwarded to someone outside the project no longer does.
A publish now records the dependency tree it actually installed — every package and version in the frontend and the backend — and the deployment view shows it. It is read from what the publish installed rather than resolved again afterwards, so it is the tree that is serving traffic. Failed and cancelled attempts are recorded too: an attempt that got as far as installing still says what it would have shipped.
The same view switches from the full tree to the difference against the previous published version of that target: added, removed, upgraded and downgraded, direct dependencies first, with the move spelled out (18.2.0 → 18.3.1). After a publish, “which package moved?” is asked far more often than “what is in this one” — a regression is usually a dependency that shifted.
Four analysis findings lived in the project scaffold itself — code nobody wrote by hand — and blocked the first publish of projects that had done nothing wrong. They are gone, and the scaffolds are now scanned with the real scanners on every change to them, so the next one cannot quietly reach a new project.
A `# nosemgrep` comment had no effect at all — the gate never saw it, so a finding you had reviewed and waved through kept blocking every publish. An accepted suppression now stops blocking the release, but stays in the Problems panel marked as suppressed. It can make a finding quiet; it can never make one disappear.
Cancel on the questionnaire card sent “decide for me and continue” as a new message and let the assistant keep building — the opposite of what the button says. It now sends nothing: the card collapses to a grey Cancelled chip and the chat input is yours again, so you can say what should happen instead.
While the assistant is writing files, the preview is broken half the time by construction — a missing import that the next write adds, a module mid-rewrite. The error counts, the tab badges and the fix button now stay quiet until the turn has ended and the re-check has landed, so what they offer to fix is the state of the finished result.
Debian published fixes for thirteen critical vulnerabilities in system libraries our images carry. Both images were rebuilt on the patched packages.
Following the rename, the addresses on the site and in the mail we send — support, sales, privacy, press — read @airbuild.ai.
When the assistant asks about colours, the question now arrives as a row of swatches — the shades it suggests for your project — next to a colour picker for anything else you have in mind. One click answers it and the colour goes straight into the design, instead of a code you have to look up and type out.
When a teammate republishes a site, the new state now shows up for everyone with access to the project rather than only for the person who started it.
OpenAI's GPT-6 Astra joins the frontier models you can choose per project. DeepSeek V4.1 Flash takes the place of V4 Pro — quicker, cheaper per task, and it can now read images you attach. Both appear in the picker with the same cost, speed and quality rating as everything else, so you can see what you are choosing before you choose it.
A deployment target always published from `main`. It now remembers which branch it builds from — pick it from a searchable list of your repository's branches, with a refresh for one you have only just pushed. Before a publish starts, that branch is brought up to date, so a colleague's push is part of the release instead of quietly missing from it.
Organisation settings now carry a default chat model. New projects in the organisation start on it, and members can still switch wherever the organisation's model rules allow it.
The secrets stored for a project — API keys, connection strings, anything you added in its settings — are now removed together with the project instead of outliving it.
A test run sat behind a plain spinner, and after a minute that is hard to tell apart from a turn that has got stuck. The card now counts along with the suite — "10 of 20 checks done", with a progress bar — and if your connection drops and comes back mid-run, it picks the count back up instead of falling back to the spinner.
The badges on the tabs, the Problems panel and the "Fix with AI" button each counted errors their own way, so the button could turn up without a badge pointing anywhere. All three read the same list now: a visible button always has somewhere to take you, and a problem you have already handed to the assistant stays marked as dealt with.
The Documentation tab re-reads the folder when you come back to it, so a page written while you were looking elsewhere is simply there. Open answer fields in the questionnaire grow with your text instead of scrolling inside a small box. The chat card header and the plan bar take up less room, leaving more of the screen for the conversation itself.
When a turn paused for want of AI Tasks, buying more did not always pick the work back up — the conversation sat and waited for another nudge. Topping up now continues the paused turn by itself, from the step it stopped on.
A scanned PDF is really a stack of photographs, and a long one could take up so much room in a conversation that the chat would not continue. Scans are now scaled down before they are sent — to the size the model reads them at anyway — and a large attachment is kept in full only where it is actually being looked at. A PDF that contains real text is never turned into pictures, so its wording stays readable to the assistant.
Every file the assistant wrote reached the preview twice, so one edit rebuilt and refreshed the page two times over. One change is one update now: the preview settles faster, and a long turn puts noticeably less load on the workspace behind it.
A Free account that has used up its starter tasks can now ask for more right on the card that stopped the work: your name, your company if you have one, and where you heard about us. A person reads it, and the request keeps a visible state of its own — "under review", with a timeframe — rather than disappearing into a queue nobody can see.
The card showed everyone the same button — "Upgrade or buy Tasks" — including plans that cannot buy extra tasks at all. Plans with add-ons now buy them in the card itself and are back at work without reloading anything; the others are pointed at the route that is genuinely open to them. A purchase also credits exactly the amount shown on the tier you picked, and the number you see afterwards comes from what was actually granted.
If the service that counts AI Tasks was briefly unreachable, the chat announced that your tasks had run out. It now says what is really going on and tries again, so a passing glitch does not send anyone to the pricing page.
Closing an account now clears the personal details out of it — e-mail address, name, sign-in identities, notification channels, anything we held to verify you — along with the content of your projects, while the billing records we are legally required to keep remain, no longer connected to a person. The same applies on every plan.
An account waiting for approval now gets an e-mail the moment it is approved, instead of the person having to try the login again to find out.
This entry originally announced that Platform Builder had been renamed to Suite Builder. That rename has not shipped: the plan is still sold and billed as Platform Builder, and nothing about what it includes has changed. We are correcting the entry instead of removing it, so the record shows both what was announced and what actually holds.
Connecting your own domain asks whether to serve it from `www` or from the bare domain, and the two looked interchangeable. They are not — the bare domain only works with certain DNS providers. Each option now describes what it needs, with your own domain filled into the text and `www` marked as the recommended one. Both fields of the verification record also have a copy button, so nothing has to be selected by hand any more.
The feature lists on the pricing cards put things in a different order for each plan and named the same thing in two ways, which made holding two plans side by side harder than it should be. They follow one order and one wording now.
Going back to an earlier version put the files back in place, but the step was not always recorded in version control — so what you saw and what your repository knew could drift apart. A saved revert is now a commit of its own, pushed along with everything else and attributed to whoever asked for it.
The workspace kept checking for project data and task usage even when nothing was running. That now happens only while something is actually being generated, which makes an open tab noticeably lighter on a laptop battery.
The backend log opens on the newest lines and follows new output as it arrives. Scroll up and it stops and holds your position while you read; one key or one click puts you back on live. Older lines load in batches as you go further back, and the errors in view are counted for you.
The credentials for a connected Git account are resolved on the server for every operation and never handed to the browser, and anything that looks like a secret is masked before it reaches a log.
DeepSeek charges different rates inside and outside its peak hours. That difference now reaches your AI Task count instead of being averaged away, so what a turn costs follows the time of day it actually ran.
The premium design option was a switch you had to find and turn on — which meant the plain default look was what most first projects got, and the look is the first thing anyone notices about a generated app. It is on for every new project now, and the switch is there to turn it off.
Google's Gemini 3.7 Flash is in the model picker in place of 3.6 Flash — the same role in the list, a newer model behind it.
Expanding the preview toolbar used to push the tab bar around. It stays put now, and the navigation keeps a clear distance from the preview.
Sign-up does not accept single-use throwaway e-mail addresses. Where that used to end in a blunt error, it now explains what happened and what to use instead.
Our answer to a support request arrived by e-mail and ended there — nothing in the product showed that a conversation was running, and following up meant digging that mail out again. Settings now has a Support tab with your own requests, the full thread on each, and a box to answer on the request itself. The e-mail we send links to the same place, so both routes lead into one conversation. Older requests show their original message, so nothing looks empty.
Reaching the end of your tasks mid-turn looked like something had broken: red card, warning sign, a single button to the pricing page. Nothing had broken — the turn stopped and everything it had done was kept. The card is quiet now. It says your work is safe, shows how far the plan got ("8 of 9 planned steps are done"), and offers three ways onward instead of one, including a box to tell us where you are stuck that reaches a person with your project already attached.
When something does not work, the follow-up question used to ask which feature you wanted next. It asks what is not working now, which is the question that actually gets you unstuck.
The home page and the Platform page now walk through all eight steps from a described idea to a published, monitored application, with a demo panel for each step, and the audience sections describe what the platform really covers for each of them — including the CRM and ERP cases people most often arrive with.
A pass over every documentation page: seven features that had never been written up are now documented, pages that still described an earlier version of the product were corrected, and the pages carry figures captured from the app itself. The workspace tab list, the project rules file, folders, taking a site offline again and the version-control features were the largest gaps.
Projects on the largest plan no longer count publications against a cap.
Running the tests produced coverage and cache files inside the project, which then turned up in your version history as changes you never made. They are ignored now, and existing projects have them taken out of version control.
Projects now have a `docs/` folder next to `frontend/` and `backend/`, and a Documentation tab that shows it. Markdown is rendered as a document, with the same labelled Source/Preview toggle the project rules window uses; other file types keep the familiar highlighted view. Because the folder lives in the project's own repository, every document is versioned together with the code it describes and travels with a clone, an export or a publish — and the assistant can see what is already documented and write into it, so a page it writes mid-turn shows up without switching tabs. Projects created before today can create the folder straight from the tab's empty state.
The free starter month came with 250 AI Tasks but kept the regular limit of 25 per day, so the 250 could only be spent across at least ten days. The first customer to try it was stopped after three of eleven steps in their first project — exactly the moment you least want to be interrupted. While the starter window is open the daily limit no longer applies: the 250 are the only ceiling. A window that is already running is lifted the next time you spend a task, and once it closes your plan's normal daily budget takes over again, unchanged.
A few paths could put German in front of people who never write German. A turn without prose of its own — a questionnaire answered purely by clicking, a pasted log, a bare "ok" — has no language to mirror, and some notices and internal examples were German-only. Airbuild now remembers the language a chat is written in and uses it for exactly those turns, as a tiebreaker that never overrides the language of an actual request. The questionnaire, its submitted answer message, the iteration-limit notice and the last German examples in our own prompts all follow the same source now.
The assistant re-runs the tests to check its own work, and every run used to render an equally prominent card — so a single message could show a green "Everything works" above five identical amber cards, and the first one printed looked like the answer. Runs that repeat the same result now collapse into one card that says how often it ran, and a card only becomes quiet history once a later run of the same scope has actually re-answered it. A backend result and an app result therefore both stay prominent, each current for what it really checked, and a run still in progress no longer greys out the last answer you had.
Fifty AI Tasks a month was enough to try Airbuild, but rarely enough to actually finish something. The Free plan now starts with 250 AI Tasks for 30 days. The month begins with your first build, not with signing up — so nothing ticks away while you are still reading. After those 30 days Free continues with its regular 50 AI Tasks per month. The daily limit of 25 AI Tasks is unchanged, and the bonus is granted once per account. Your remaining tasks and the days left in the window are shown in the AI Tasks panel in the sidebar and in Settings > Subscription.
Every account on a paid plan received 250 AI Tasks as a one-off gift. They land in your top-up balance, which means they never expire: they carry over from month to month and survive a plan change. You can see them in the AI Tasks panel — the total includes your monthly allowance plus everything in your top-up balance.
The usage bar under Settings > Subscription used to compare your balance against a hardcoded plan number, which would have read "120 of 50 used" during a starter month. Balance, limit and progress now all come from the same server-side figure, so the number in the sidebar, the bar in Settings and the pricing page can no longer disagree.
Two limits apply on the Free plan — a daily bucket and a monthly allowance — and only one of them was ever reported. In a starter month you therefore saw something like "22.7" next to a badge saying "250 this month", with no way to tell what was left of the 250. Every free display now names both, "22.7 of 25 today" and "247.7 of 250 this month", from one shared calculation. The same change ends "227 of 250 used" showing up on day one, and shows real usage for the current period on paid plans, where top-ups sit on top of the allowance.
The picker had a fixed height from a time with fewer providers, so adding OpenAI pushed its group exactly one row below the fold — and since nothing on screen hinted there was more to scroll to, the newest provider looked like it did not exist. The panel now sizes itself to the number of providers, and it decides whether to open upwards or downwards from that real height.
Now that values are masked everywhere, copying was the only way to check one. Every value has an eye toggle next to Copy — on the Environment page and in the deployment-target editor. Values are hidden again after a reload, a target without its own override reveals the inherited value it will actually use, and a row you have just added starts visible, because there is nothing stored yet to protect.
The target you publish to by default is now called "Production" instead of "Default" — a name for the thing itself, not for its role in our data model — and the redundant "default" chip next to it is gone. Settings moved to the front of a target's tabs, since address, hosting and SEO live there and nothing else works before those are set. In the environment table, the Development column now sits right next to the variable name, because that is the value you read and edit while building.
Routing for published sites was decided from the state of the last publish attempt rather than from the version actually being served. A cancelled attempt next to a healthy live site therefore read as "offline" and its routes were removed — that took one custom domain down on 18 August — and while a re-publish was running, a site was skipped entirely. Routing now follows the version that is really live, so a site that has served traffic keeps its address until it is genuinely taken down.
We read through what actually went wrong in the 45 turns that crashed in the week to 19 August, and four of those causes now have their own message. The most important one: a provider account out of credits arrives as a rate-limit error, so those users were told to wait a moment and continue — advice that could never work. That case now says what it is and suggests switching models. Two of the others are our fault, a request we built wrong for that model and a chat history we failed to resume, and the copy says so plainly instead of implying you did something wrong: the first is solved by another model, the second simply by continuing.
Nine turns between 12 and 18 August ended in the anonymous "the generation stopped unexpectedly". The cause sat in how we hand a model back its own reasoning when a conversation continues: text segments were merged, which moved reasoning blocks out of the arrangement the model had produced — and the provider rejects exactly that. The arrangement is now recorded as it is emitted and replayed unchanged.
Not every AI model can look at pictures. If you attach an image while a text-only model is selected, the composer now shows a friendly amber hint before you send. Nothing is blocked — the file is still added to your project, only the image analysis is skipped — and switching to a model that can see images is one click away. The list of which model sees what now comes from the backend, so the hint can never fall out of date.
The Environment page and the deployment-target editor used to show variable values in clear text unless a variable was flagged as sensitive. Every value is now masked regardless of that flag, and a Copy button next to each one is the only way to read it — so nothing sensitive ends up on screen during a screen share or a screenshot.
Error cards in the chat were translated while they were live, but reverted to English the moment you reloaded the page. They are now translated whenever they are shown, from the history too, and they name the concrete cause and the file involved. A skipped image analysis also no longer appears as a red failure with a "Try again" button — it is a neutral blue notice, because nothing actually failed.
A JPEG exported as "logo.png" — something design tools do routinely — used to be announced to the AI with the wrong format, and the whole request was rejected. Uploads are now named after what the file actually contains, so a mislabelled export is silently corrected. Files that are not images at all are refused right away with a clear message instead of failing later.
Cancelling a new publish of an already-live site could remove the routing its previous version was still serving — the site went dark until the next repair pass. A publication that has served real traffic now keeps its address no matter how a later publish ends, and routing is finalised on every publish outcome instead of only some. Re-publishing also reliably re-claims a custom domain that had lost its route.
When the preview could not start, the panel always claimed the dependencies failed to install — even when they installed fine and the dev server was the problem. It now reports the stage that really failed, with the original message, and the assistant gets pointed at the right thing instead of hunting through dependencies that are already fine.
If a dependency could not be resolved, the workspace was still reported as running with a working preview — so you saw a dead frame with no explanation while everything claimed to be healthy. The preview state is now recorded honestly by every path that starts or restarts it, so the error panel appears immediately and clears itself as soon as the preview is back.
Between 15 and 17 August, publishing a Flex-hosted app could fail with an internal quota error. Two leaks were behind it: re-publishing left the previous API connection behind instead of reusing it, and unpublishing removed things in the wrong order and orphaned them. Both are fixed, the accumulated leftovers were swept, and lookups now page correctly so nothing is missed once you have many apps.
Which AI models your members may use was previously only applied by the model picker in the workspace. It is now enforced by the services themselves — when a project is created and whenever its model is changed — so the setting holds even outside the interface. Nothing changes for you if you were already picking allowed models.
Project rules now have their own entry in the mobile project menu, right where they sit on the desktop. Collaborators who are not the project owner get a clear read-only view — they can read the rules and the preview, but the editor, the Save button and the shortcut are hidden instead of failing on save.
Features that are not live yet used to carry a single generic "coming soon" flag. Each one now names its own quarter, so you can see at a glance what lands this quarter and what comes later — Agent Flow and Skills are marked for Q4, the remaining upcoming features for Q3.
xaio.dev (now airbuild.ai) got a substantial update: a comparison hub that puts Airbuild side by side with other builders across deployment, migration and business-solution dimensions, a complete press kit with official logos and product screenshots, a public blog with ten new posts, and a dedicated page on Airbuild as an AI coding tool. Colours, headlines and the German wording were reworked throughout, and mobile headlines finally wrap the way they should.
A freshly created project used to offer a grey Publish button and no explanation. While a deployment target is incomplete, the main action is now "FINISH SETUP" above a list of exactly what is still needed, and it opens that target's settings with the required fields marked. Saving hands you straight back to the publish dialog, ready to go. The tab checkmarks, the blocker list and the Publish button are all driven from the same source now, so they cannot disagree with each other.
A target could previously reach the publisher half-configured: unnamed, with hosting never chosen (a web app then quietly shipped without its backend), or with SEO switched on and no title (the publisher invented one and sent it to search engines). These are now required — checked both when you save the settings and when you publish, so anything you can save is publishable. The first target of a project may still stay unnamed; it simply becomes "Default".
The publish request carried the dialog's own defaults rather than the target's settings, so a target set to Private L published as Cloud and offered the wrong tier when ordering hosting. Hosting type, resource size and database now all come from the target itself. If publishing is refused because hosting is missing, the hosting order opens instead of a dead error message.
Publishing a paid web app could be refused with "Active hosting subscription required" while the subscription was billing perfectly normally — the subscription was never linked to the deployment target that the check looks at. A project's default target now adopts its hosting subscription on the first publish, and existing subscriptions were linked up retroactively. No charge, no status change, nothing to do on your side.
Generated apps now declare APP_NAME as a real variable, pre-filled with your project name. It holds the application name alone, so it can be reused wherever the app names itself — emails, page titles, PDF headers — and each whitelabel target can override it to name itself. It also stops the publish preflight from nagging you about a variable that was already set sensibly.
Copying dependency folders, cleaning up workspaces and searching through a project used to run on the same thread that also relays your live chat — so a big file operation could freeze chat and preview updates for everyone on the same server. Those operations now run out of the way, and file searches skip dependency folders entirely instead of walking through them.
A project used to own no deployment target until its very first publish, so the Deployments page opened empty and "Edit in Settings" pointed at something that did not exist. Every project — new, imported or cloned — now gets a default target the moment it is created, and existing projects were given one too. It is inactive and holds no address: nothing is deployed, nothing is charged, and no *.xaio.app name is reserved for a project that may never be published. "Enable SEO" now defaults to on for new targets.
A first publish could dead-end with a greyed-out button and no reason given — most often because the suggested address was already taken. The action bar now names the actual blocker (no address, address taken, invalid custom domain) and links straight into the target's settings, where the address is pre-filled from your project name, marked as required, and clearly labelled as only claimed on save. The Domain tab also marks a not-yet-claimed address as "(suggested)" instead of presenting a guess as a configured domain.
Opening a new chat in an existing project could re-send the project's original description, and the assistant set about rebuilding an app that was already finished. The original request is now only ever auto-sent in a project's very first chat. Project details also update live and are re-read on every reconnect, so the workspace no longer works from a stale copy. As a side effect, the context counter next to Compact updates immediately instead of after a reload.
The chat used to accept only a fixed list of file types. You can now attach anything — documents, archives, data files — with a preview card that recognises the type and a matching icon. The upload limits and safety checks stay in place.
Requesting a Platform Builder trial is now a two-step form: name, company and business email first, then four optional questions you can skip entirely. Once a trial is running, it is visible where it matters — the plan card reads "Current · Trial" with the end date, and the billing tab shows trial details (end date, seats in use, remaining AI tasks and a shortcut to invite members) instead of a seat-purchase form you should not need during a trial.
Several chat messages claimed a cause the system could not actually know — "AI service temporarily unavailable" appeared for any unclassified problem, and an interrupted turn was described as a deployment even when it was a real failure. The wording now states what happened, that your work is saved, and how to continue, without inventing a reason. Three message types that were still shown in raw English to German users are now translated.
Models outside your organization's allowlist used to appear greyed out with a lock icon. They are now hidden entirely, and providers whose models are all blocked disappear from the sidebar. Saving an empty allowlist — which would have locked the whole organization out of every model — is refused, in the settings and on the server.
Opening a project without access fired off every request the project needs anyway — chats, git, integrations, workspace — and each one failed, stacking an error toast on top of the not-found screen. Access is now decided first, and nothing project-specific loads until it is granted.
The Organization tab in your personal settings now goes directly to your organization settings when one exists, instead of stopping at an overview page first — and creating a new organization lands you in its settings too. The back arrow in the organization header returns to your personal settings rather than jumping to the project list.
Platform Builder is on general sale, and the plan card now carries a split "Upgrade | Trial" action. The trial request works logged out too, validates business email addresses inline, and a granted trial provisions a ready-to-use organization with its own AI-task pool — no credit card, no Stripe subscription. Trial organizations can restrict invitations to a single email domain, and unknown invitees from that domain are added automatically with their own access mail.
If you bought a top-up on a paid plan and later downgraded or cancelled to Free, those tasks were stranded: visible in your balance but impossible to spend, with no refund and no rollover. The Free plan now spends your free daily allowance first and then your top-up for the rest, exactly like the paid plans do — and the balance shown is correct again.
The model badge on a message used to be a guess, defaulting to Sonnet when the answer had in fact come from somewhere else. Each turn now records the model that really ran, along with the mode it was sent in, so what you read on a message is what produced it.
Every message used to be written by rewriting the entire conversation file, and the whole history was re-read to verify a single new line — the cost grew with every message in a long chat. History is now appended, per-turn details live in their own file, and long conversations no longer weigh the service down. Only one process may write a chat at a time, so a restart mid-turn can no longer scramble it.
The assistant can now sketch quick, self-contained UI prototypes while you explore an idea. They live in the new Prototypes tab, where you can view each one and click "Build this" to turn a sketch into a real change — the request lands in the chat so you can edit it before sending.
Give the assistant a URL and it captures the website for you — pages and assets land in your project via the Airbuild Browser integration, with a live progress card in the chat. Great for rebuilding or migrating an existing site.
When the pre-release checks flag something, you now get a proper review dialog: pick exactly which findings to address, hit "Fix with AI", and resume the held publish once you are satisfied — failures link straight to the QA details.
Kimi K3 joins the model picker and can be selected for your projects.
Published sites get a new security setting: decide which external script sources your site may load — handy for analytics or embeds — without weakening the strict defaults for everything else.
The publish dialog is slimmer: pick a target, see a summary and the pre-publish checks. API endpoints you intentionally leave open can be marked as public by design and reviewed in one place, and hosting options are shown as clear tier cards with details on demand.
Publishing now reports accurate progress from start to finish, failed deploys no longer get stuck halfway, and each project runs exactly one publish at a time — no more competing deployments. Cancelled publishes end cleanly, and Airbuild warns you when you are about to publish from a branch other than main.
The chat now shows which AI model generated each answer, tells you clearly when a long task pauses so you can continue it, and the file browser refreshes live while the AI changes files. Images added to the media library appear as thumbnail cards right in the conversation — and a flaky connection no longer leaves you with an empty chat: history loads retry automatically.
Opening a published site from the project menu did nothing on mobile browsers — the tab was silently blocked. Fixed: the published site now opens reliably, on iOS Safari too.
A broad hardening pass: published-app builds now run in fully isolated environments, and numerous internal checks around publishing, sharing and integrations were tightened. No action needed on your side.
Organization owners can now set a monthly AI-task budget per member, not just for the organization as a whole. Usage is tracked against each member's own budget, so one member cannot quietly consume the whole shared pool.
The org-wide rules you define in your organization settings are now injected into every chat in your organization's projects, alongside the project's own rules — so a house style or a technical constraint written once actually applies everywhere. The model allowlist is applied by the model picker at the same time.
The deployment-target manager, target settings, hosting settings and the Environment page were hardcoded English. They are fully translated now (EN/DE). The visual SEO image pickers are back too: a social-image preview in the right proportions with Replace and Remove, plus a square icon tile, instead of plain URL fields.
The assistant used to stop after a fixed number of working steps, which cut off genuinely large tasks halfway. That ceiling has been raised substantially, so a big refactor or a multi-file feature can finish in one go.
Switching between deployment targets kept the previous target's label, address, hosting and SEO values on screen — the settings form never re-read the new target. Each target now loads its own values when you switch.
Platform Builder — organizations with shared teams, org-wide projects and simple per-seat billing — is now on sale for everyone.
Organizations get their own guardrails: define org-wide rules (AGENTS.md) that the assistant follows in every project of your organization, and choose which AI models may be used by your members. Both live in the new organization settings.
Airbuild now labels AI clearly everywhere: the chat notes that responses are AI-generated, assistant messages carry an Airbuild AI label, and Agent Flow marks requests that come from AI agents. Published sites and generated images additionally carry an invisible, machine-readable marker for AI-generated content — in line with the EU AI Act.
The AI model picker got a full redesign: search, filter chips and a provider sidebar help you find the right model, and every model is described in plain language instead of developer jargon. The menu also never gets cut off anymore, even in tight layouts.
Actions like deleting a table row, disconnecting GitHub or closing unsaved project rules now ask with a polished in-app dialog instead of the plain browser popup — with clear buttons and matching design in light and dark mode.
The preview now recovers on its own when your workspace stopped in the background, and the file browsers refresh live the moment the assistant writes or changes files.
Cloning a project now creates a genuinely clean copy: code, database setup and integrations are set up fresh for the new project, and you land in it right away with a live status while everything gets ready.
The publish dialog scrolls properly in Firefox, the progress bar no longer jumps backwards when you confirm an advisory note, and your hosting subscription status refreshes reliably after publishing or unpublishing.
You can now build with the latest model generation: Claude Opus 5, Sonnet 5 and Fable 5, Google Gemini 3.6 Flash, and OpenAI GPT-5.6 in the Sol, Terra and Luna variants. The 5-series models use adaptive thinking and reason exactly as much as your task needs.
The assistant can now look into your version history: ask what changed, when and by whom, and you get friendly cards instead of raw output. Changes the assistant commits are attributed to the teammate who asked for them.
If a release is stopped by a quality check, Airbuild now switches you straight to the QA tab with the failure details — no more hunting for what went wrong.
Every published app is now built in its own fully isolated environment, and the pre-publish security checks are strictly fail-safe: if a check cannot run, the release stops instead of slipping through.
Platform Builder now has straightforward per-seat pricing at 150 € per month — consistent on the website, in checkout and on invoices. Invoices go automatically to your organization’s billing contact.
The project creation form is now one and the same across every entry point, and the plan-my-website button activates as soon as you have typed three characters.
The address bar now remembers which project tab and monitoring section you are on — copy the URL and your teammate lands exactly where you are.
Findings from the publish checks now live in Monitoring under Problems. After a fix-with-AI run the scanner automatically re-checks and shows only what is left — resolved findings clear on their own.
Confirming a two-factor code no longer fails with a network error in some situations — signing in with 2FA is dependable again.
Login and signup on the website are now fully bilingual and keep you in your language all the way from the pricing page to your account.
Every project now has a Rules tab: write down conventions, dos and don’ts or anything the assistant should always respect — the rules apply to every AI turn. With markdown preview, and you can also simply ask the assistant to remember a rule for you.
Files you attach in chat start uploading immediately, with visible progress per file — hitting send no longer makes you wait.
Deployment history now shows the analysis and security checks of every publish attempt, with findings right in the view and the full report one click away. Each attempt also gets its own version number, so failed tries are clearly told apart.
When a chat message fails, the error card now has a proper try-again button.
Platform Builder brings organizations to Airbuild: switch between your personal space and your organization, keep projects and teams org-wide and manage members centrally.
xaio.dev now hosts the full documentation and this changelog — including Ask Airbuild, a small assistant that answers questions about the docs right on the page.
Conversations can now be archived and brought back anytime — archiving replaces clearing, so no chat history is ever lost by accident.
Pre-publish tests now run against a frozen snapshot of exactly the code that ships, with their own fresh test database and warmed caches — noticeably faster and immune to changes you make while the release runs. The publish view also shows your live version separately from the latest attempt.
Queued follow-up messages now sync in real time for every collaborator in the chat.
AI-task top-up packs now charge exactly the volume discount that is advertised.
If a frontend environment variable looks like a secret API key, publishing now blocks it before it could end up publicly visible in your app.
Long backend logs now load in pages with a load-older button instead of all at once — even big logs open immediately.
The assistant can now write noticeably longer responses and bigger files in a single turn — output limits were raised to each model’s true maximum.
Analysis checks during publishing can no longer hang, and cancelling a release now also stops a check that is mid-run.
Automated testing is now switched on for all plans: the assistant runs your project’s tests as it works and reports test coverage by default, so you can see how well your app is protected.
When the chat is centered, the navigation toolbar steps out of the way and the project name is neatly centered.
When you publish, the quality panel now suggests the checks we recommend for a production release and pre-selects them the first time. Expand any check to see what it does, then publish — the checks you keep selected are switched on and run as part of the release.
Publishing now shows a single progress bar that climbs smoothly from 0 to 100%. Every step — endpoint check, code analysis, dependency scan, tests, build, upload and verification — has its own portion of the bar, long steps keep moving live, and the bar never jumps backwards, including the handover from analysis to deploy. A failed step you retry no longer gets stuck.
New projects now start with our recommended integrations — like code-quality and security scanning — already enabled, so the assistant uses them from the very first turn without you having to switch anything on.
On mobile, the project menu now stays within the visible screen and scrolls all the way to the last item, so options like Publish and Open Published are always reachable.
The Logs / Files / API Docs panel now adapts to narrow layouts and phones: pick a file and its contents open full-width and readable, the tabs no longer wrap, and the status bar stays tidy.
While the assistant is planning, the individual planning steps no longer clutter the chat as separate cards — the Build Plan card is now the single place that shows plan progress, in both chat and project view.
You can now open version control and team management straight from the project menu on mobile, so you can review changes and manage collaborators on the go. Access follows the same rules as on desktop.
The "What do you want to build?" screen now lays out cleanly on small screens: the model picker, premium design and submit buttons stay on one tidy row without overlapping or getting cut off.
On mobile, the Publish action now only appears for owners and admins, matching desktop — so collaborators without publish rights no longer see it.
When the pre-publish check flags publicly accessible endpoints, you can now select exactly which ones to address and hand them straight to "Fix with AI".
Chrome's password autofill popup no longer appears over the mini-game shown while your app is being generated.
As part of 1.0 we completed an internal, platform-wide security review and strengthened authentication, fine-grained per-project access controls and tenant isolation across every backend service, so your projects and data stay isolated to you and your team.
Large screenshots and chats containing many images are now resized automatically, so requests no longer fail — the assistant keeps responding instead of getting stuck.
The dynamic security scan that runs before publishing can now sign in to your app, so it also checks authenticated pages and reports any findings in the problems panel.
Beta badges have been removed from the sidebar, settings and the mobile project list for a cleaner, more finished look.
Before publishing, you now get an advisory security review of your app that you can look over and acknowledge, so you go live with a clear picture of any risks.
Test runs now show their progress right in the chat while you work in project mode, so you can follow along as each test completes.
Published-site routing now self-heals if it drifts, and hosting period-end dates are kept correct — so your live sites stay reachable and your billing dates stay accurate.
Fixed usage tracking so session activity and per-model usage are recorded reliably, keeping your usage figures accurate.
When you delete a project, its managed Git repository is now removed as well, so nothing is left behind.
Fixed a case where a workspace slot could stay reserved after use and reduce available capacity — so workspaces start reliably when you open a project.
Before publishing an app with a backend, Airbuild now shows which API endpoints will be publicly accessible and asks you to confirm, so nothing goes live open by accident.
The publish QA tab now orders checks by the sequence they run in and lets you expand test results inline, mirroring the Tests tab — so it's easier to see what passed and dig into the details.
Apps you generate are now built on React 19, keeping your projects on the latest stable version.
Automated tests now run against a dedicated, isolated database, so test runs are reliable and repeatable and never touch your real data.
Publishing has been rebuilt around a new QA tab: watch every quality, test and security gate run in real time, in the exact order they execute, each with a per-test checklist and a clear pass/fail state. When a gate fails, its findings hand off straight to the Problems panel so you can fix them with AI and publish again — no more guessing why a publish was blocked.
On-demand security scans now list all known vulnerabilities, including ones without a fix yet, so you get the complete picture.
After a successful publish you now get a "Publish checks" panel showing which quality, test and security checks ran and passed — so you have proof your app went live in good shape.
The Tests tab now has its own coverage view, so you can quickly see how much of your app is covered by tests without digging through reports.
Before an app with a backend goes live, its API is automatically validated against common best practices, catching sloppy or inconsistent endpoints early.
Your app can now be scanned while it's running, and any security findings appear right in the Problems panel — so you can spot and fix real-world vulnerabilities in one place.
You can now switch on a deeper check that inspects the packages your app depends on for known vulnerabilities before it goes live, giving you extra peace of mind.
Testing now has a clearer name and a simple toggle in settings to run your tests as a gate before each publish, so broken code never reaches your users unnoticed.
Mistral Medium is now available in the chat model picker, giving you another option to match the model to your task.
Frontend and backend tests can now run automatically before publishing, with a clear "running tests" step. Publishing is only blocked when there's a real test failure, so your users see only working versions.
The cap on the number of published sites has been removed on paid plans, so you can ship as many apps as your work requires.
While the AI is writing files, the chat now shows a clearer placeholder so you can tell it's actively working rather than stalled.
A unified Problems panel now shows live code diagnostics as you build, each with a one-click "Fix with AI" button so you can clear issues without leaving your flow.
You can now trigger automated code quality and security checks on demand right from chat, with findings neatly grouped by the type of check so they're easy to work through.
The Tests tab now shows coverage, groups your test cards by file, and links to a full report — making it much easier to see where your app is well tested.
The Tests tab has been rebuilt into a full test cockpit: run tests on demand, see per-test results with filters, browse run history, explore a file-to-test-case tree, read plain-language run rows, export results, and watch test runs appear as a live card in chat — all with friendlier, jargon-free settings.
When a custom domain hits a fixable setup problem, it now self-heals and retries instead of failing outright — so more domains connect without you lifting a finger.
Collaborators can now use version control on projects that are bound to a connected Git repository, making team work on those projects much smoother.
If you interrupt the AI mid-turn, your conversation history no longer shows duplicated messages, so the chat stays clean and accurate.
The Tests tab now lets you provision, run and gate automated tests for your app, so you can build in confidence that things keep working as you change them.
While plan details are still loading, plan cards now show a subtle shimmer placeholder instead of appearing empty, so the pricing view always feels responsive.
Payment notifications that fail are now retried automatically, and monthly plan allowances are granted correctly on renewal — so your billing and included usage stay accurate.
Cloud hosting is now offered based on whether your project actually uses a backend, not just its project type. So you only see — and pay for — hosting that fits what you're building.
When something goes wrong in chat, the message now shows up in your language instead of a generic English notice, so it's easier to understand what happened.
The project header menu can now scroll on smaller screens, so every option stays reachable instead of getting cut off at the bottom.
Generated apps can now be set so that only admins create accounts, instead of letting anyone sign up — perfect for internal tools and invite-only apps.
Mistral Large 3 and GPT-5.5/5.4 now accept image input too, so you can share screenshots and mockups with more of your favorite models.
During publishing, the "Fix with AI" button is disabled while the AI is already working and re-offered after a failed retry — so you never fire off overlapping fixes.
The chat status line no longer leaks internal control markers, so you only see clear, human-readable progress.
Who a collaborator is and who can manage them is now determined on the server, with management limited to the project owner — so team access is always consistent and correct.
The model picker now includes OpenAI GPT-5.5 and GPT-5.4 as well as Mistral Large, giving you more options to match each task to the model you prefer.
When a code-quality scan flags issues before publishing, the "Fix with AI" button now receives the full list at once, so more problems get cleaned up in a single pass.
Switching branches now shows a reliable overlay while it loads, and offers a manual reload button if a switch takes longer than expected — so you're never left wondering whether the preview is stuck.
Additional AI models are now available in chat, so you can pick the one that best fits the way you like to build.
The live status feed is back during project edits, and reloading a conversation no longer shows duplicate images — your chat looks exactly as it should.
If a custom domain is missing its security certificate, it now shows as failed instead of staying stuck in "pending", so you know to take action.
When a preview is just warming up, it now retries automatically instead of getting stuck, so your app appears as soon as it's ready.
If installing dependencies fails, your workspace now stays up and shows a clear panel with a one-click "Fix with AI" button to resolve it, instead of leaving the preview broken.
The monitoring view now has a "Fix with AI" button next to production errors, so you can jump straight from spotting a problem to fixing it.
When the AI changes several files at once, the chat now labels each action by what it actually did instead of always saying "create", so you can see exactly what happened.
The "Copy from" option in the environment panel now adds only the variables you don't already have, and the list refreshes when you switch tabs — so you never accidentally replace existing values.
Projects connected via GitHub login now get a clear nudge to install the GitHub app, unlocking the full set of repository features.
Sending a message or attachment now survives a brief loss of connection: you'll see a "Reconnecting…" indicator and it retries automatically, so your message isn't lost.
A new "Add organization" button lets you install the GitHub app for an entire organization in one step.
When a paid subscription ends, your account now moves to the Free plan automatically, so your projects keep working without an awkward in-between state.
Alerts about high usage now include the project name, so you immediately know which project to check.
If an edit gets interrupted, it's now finalized and recovered properly instead of leaving your chat in an unclear state — and older messages load as you scroll up.
When the AI edits several files at once, the chat now shows the matched code snippet for each file, so you can see precisely what changed where.
When pulling changes from your repository raises a warning, it now appears directly in the interface instead of passing by silently.
The Design Mode prompt bar and the custom-domain status badge are now fully translated, so the whole interface reads naturally in your language.
Emails about a specific project now include the project name in the subject line, making them easier to spot in your inbox.
Messages now save reliably even during very active editing sessions, so you won't see lost messages or sudden errors in the conversation.
Published sites now include per-page structured data and a keywords field — which you can fill in yourself or generate with AI — to help search engines understand your content. A new keywords chip field in the SEO tab makes adding them quick.
The preview now offers an orientation submenu for tablet and mobile, so you can check how your app looks in both portrait and landscape.
Image files now show as thumbnails in the file browser, so you can spot the right asset at a glance.
When the AI changes several files in one step, the chat now lists every affected file, so you always know what was touched.
The version control panel now lets you commit changes manually and even suggests a commit message with AI. The commit list refreshes right after, so you always see the latest state.
If a turn gets interrupted, the chat now shows the changes it made so far along with a clear "committing…" indicator, so nothing feels like it vanished.
As a project owner you can now always open and work in your project; only deploying is gated by your plan. Hosting emails also include the project name and direct links.
Questionnaire options with long labels now wrap to multiple lines, so the full text is always readable.
A new hosting settings page lets you reactivate hosting from the grace period and switch between cloud and private hosting. If a payment is past due, you'll now see a clear banner so you can resolve it quickly.
Sending images to the AI now works more reliably across more models, and the chat recovers gracefully if a tool response comes back malformed.
A new shared cloud hosting tier lets you host your app for €5/month, with a clear card in the publish flow showing the price and details, and a simple subscribe step.
You can now publish and connect a custom domain in a single step, with guidance for root/apex domains (like example.com) including automatic www redirect. Remove the domain and your site cleanly reverts to its xaio.app address.
New hosting now starts with a real 6-hour free test phase so you can try it out, with a clear email when the trial is ending. After that, a daily lifecycle keeps things tidy with a migration deadline and a grace period before shutdown.
You can now register your interest in the upcoming Platform Builder plan directly from the pricing page, so we can let you know when it's available.
When generating a preview image for your published site, it now correctly uses your uploaded social image and dismisses any cookie banner, so the result looks clean and on-brand.
Your AI task usage is now counted accurately per project and per model, so the numbers you see match what you actually used.
The live preview now shows your active custom domain, and sites that use only a custom domain get their own monitoring panel so you can keep an eye on them.
Your SEO settings now reliably persist when you publish, and project owners can decide which team members are allowed to publish.
Fixed a problem where the project detail and cost views could fail to load. They now open reliably with accurate per-project totals.
Fixed a problem where edits weren't being committed and pushed to your repository automatically. Your work is now saved reliably, and inviting a collaborator to preview a project works smoothly again.
Your monthly AI tasks are now reliably granted when your subscription renews, and your subscription end date stays in sync. Affected accounts have been topped up retroactively.
Right after you publish, your site now serves the fresh version immediately instead of occasionally showing an outdated page.
Website Builder Pro, Web App Builder and Platform Builder now include unlimited custom domains (Website Builder One keeps one). Connect as many of your own domains as you need — no per-domain limit.
Every plan — including Free — can now publish to unlimited yourname.xaio.app subdomains. Free sites carry a "Built with Airbuild" badge, which paid plans can remove.
You can now switch a project from a Web App back to a simple Website. The change is blocked while a backend is still active, so you never lose connected data by accident.
The Platform Builder plan is now clearly labeled 'Coming soon' with its planned availability date. You can still explore the card, and choosing it opens a short note instead of starting checkout.
Plan cards now spell out unlimited custom and xaio.app domains where they apply, and the free plan clearly shows unlimited xaio.app domains plus the 'Built with Airbuild' badge so you know exactly what's included.
You can now create web apps on any plan. Projects switch automatically between Website and Web App when you add or remove a backend integration, so you no longer manage the type by hand. Backend hosting and publishing remain part of the higher plans.
Backend and custom-domain options now stay visible but show a clear upgrade badge if your plan can't use them yet, with one click to the pricing page. The redundant upgrade banner that always sat at the top of the Backend tab is gone, and you'll only see the prompt when you try to publish.
After the AI edited your files, the follow-up suggestion chips stopped appearing. They now show up again to help you decide what to do next.
On the free plan your available tasks now correctly reflect the monthly reset, so you no longer see a balance of zero once your period has rolled over.
Fixed a problem where your custom global styles could vanish when the app was regenerated, leaving the preview unstyled. Your styling now stays applied reliably.
Adding a backend or database now requires a plan that supports backend hosting. If your plan doesn't, the chat clearly asks you to upgrade instead of failing silently.
When you select an element in Design Mode, the chat now always shows which element you targeted, and clicking the reference opens the right file instead of getting stuck loading.
When you generate a social/preview image with AI, it now reflects your app's actual content and brand colors instead of a generic look. The result matches what your app is really about.
The preview toolbar buttons (design mode, console, reload, open in new tab) now reliably appear once your preview is rendered, instead of occasionally staying hidden for the whole session.
Fixed an issue where upgrading, downgrading, or cancelling a subscription could silently fail to update your plan and features. Your current plan now always reflects your latest billing change.
Inviting teammates to a project and connecting your own external Git repository are now available from the Pro plan upward. If you move to a lower plan, collaborators and external repo connections are removed automatically.
Apps with a backend can now be hosted on the Web App Builder and Platform plans. If you leave Platform but keep backend hosting, larger instances are automatically scaled down to the Business size that Web App Builder includes.
Plan cards now show accurate, up-to-date AI task allowances, domain limits and team-member counts straight from your plan, and we removed features that weren't actually offered. Hosting tiers are also shown more clearly per plan.
The top tab bar has a cleaner look with labels appearing on hover and smooth animations. The environment, integrations and database views are now fully translated in English and German, and database tables refresh automatically after the chat makes changes.
When you started a new project, the first message sometimes triggered two replies and two questionnaires. New projects now start with a single, clean response.
The usage tooltip now shows the reasoning tokens actually used for your request instead of a fixed budget, and correctly labels the model provider.
The interactive plan and questionnaire cards sometimes failed to appear in the chat. They now render reliably so you can review plans and answer questions as intended.
Across all models, the AI reasons through your request before responding, leading to more reliable plans and edits. You can see when it was thinking right in the message details.
If your app's backend doesn't come up during a deploy, you now see a concrete reason (out of memory, crash, stuck) instead of a vague timeout, and the deploy stops faster rather than waiting it out.
A glitch while refreshing the CDN cache could abort a publish near the end even though your site was already live. Now your publish succeeds and you simply get a heads-up that the cache may take a moment to update.
Starting a workspace now shows a calm progress indicator with a 'Workspace starting' cover and an elapsed / ~15s countdown, with a gentle 'taking longer than usual' note past 15 seconds. Previews fade in smoothly and reopening an already-running workspace no longer flashes a loading screen.
Asking @help in the chat now reliably returns documentation answers, instead of coming back empty in some cases.
You can now choose DeepSeek V4 Pro as a model for your chats, giving you another option for building and editing your app.
The model picker now includes DeepSeek V4 Pro and Opus 4.8, and shows the exact model variant you've selected (for example Sonnet 4.6 or Gemini 3.5 Flash) rather than just the brand.
You now receive helpful product announcements, like new features and important notices, as polished emails or directly inside the app.
When you picked a specific Claude model, the chat sometimes still ran on a different one. Your choice is now always used and billed correctly.
All transactional emails, from sign-up and invites to invoices and notifications, now share a clean, consistent design with a bold hero banner and clear call-to-action buttons.
A short, momentary hiccup while your preview was starting up could leave it stuck on a "server not responding" error. Your preview now rides out these brief blips and stays available.
Use plan mode to have the AI map out the work before making any changes, or ask mode to get answers about your project without touching your files. Both keep the AI strictly read-only.
Important announcements now appear in a banner at the top of your workspace. High-priority messages stand out, longer messages can be scrolled, and you can mark them as read or dismiss them.
Questions no longer add a redundant "Decide for me" option. Just leave a question blank and we'll decide for you.
The branch list could appear empty for collaborators. It now always loads the current branches from your repository.
Removing a variable in the Environment tab now saves correctly, instead of the deleted entry quietly reappearing a moment later.
If a published app starts crash-looping and is automatically stopped to protect it, you now receive an email so you know to take a look.
Images attached to a message you queue while the assistant is still working are no longer dropped — they're now sent along with the message.
If your Git branches can't be fetched (for example due to a credential problem), you now see a clear error instead of a silently empty branch list.
A published web app can now be exported from the publish dialog as a single ZIP: the app's database as an SQL dump — structure and rows, ready to restore with psql — together with the frontend that is deployed for it. On Power hosting the archive also carries the backend bundle the app runs from. Files that your app's users upload live in separate storage and are not part of this archive. Available for apps on Flex and Power hosting.
When the assistant asks a set of questions, you can now cancel out of it. Cancelling lets the assistant decide for you and continue, so you're never stuck answering.
The Edit, Copy and other actions on your messages are now always visible, including while a response is being generated. Copy stays usable, while edit is briefly unavailable mid-generation with a clear tooltip.
Rate any AI response with a thumbs up or down, and optionally leave a short comment. Your feedback helps us improve answer quality over time.
Direct messages from teammates now appear in a clearer, stacked layout — sender, the related project (one click to open) and a relative timestamp ("just now", "2 hours ago"). Links are clickable and open in a new tab, line breaks are preserved, and long messages scroll inside the banner.
If one of your published apps repeatedly crashes on startup, Airbuild now detects the crash loop, stops restarting it, and emails you so you can step in and fix it.
Mention a teammate with @name in a project chat to send them a private direct message. They receive a notification banner and can reply inline. The @-command menu was redesigned for quicker access to mentions and commands.
A new Ask mode lets you ask questions about your project without triggering any code changes — ideal for understanding your app or planning next steps. The separate Plan mode has been removed; build plans still appear automatically when the AI needs your confirmation.
Purchase dedicated private hosting for an individual published app, with a one-time 6-hour free test phase per project so you can try it before buying. Existing Power apps can be migrated to the new hosting on a clear timeline.
A slim message banner now sits above the workspace and project list, surfacing important notices — hosting reminders, team messages — one at a time. Acknowledge or dismiss them with a click.
Your interface language (German or English) is now saved to your account and applied everywhere, including all emails and server messages.
When unpublishing an app you can now choose to keep its database, and download the export first — the database as an SQL dump plus the deployed frontend, in one ZIP.
Claude Opus 4.8, the most capable Claude model, is now selectable in the chat model picker.
When you have a GitHub repository connected, the automatic commits made by the AI are now pushed to your repo, keeping your history in sync.
Reopening a project whose workspace is still warm is now noticeably faster, skipping unnecessary startup steps.
Publishing now streams its progress live over a websocket instead of polling, so the publish dialog and button update instantly and reliably.
The chat handles large PDFs more reliably now, with dedicated read modes and protection against truncating big documents.
The initial project questionnaire now always offers a free-text answer and drops the "decide for me" option for a clearer flow.
Fixed a Stripe webhook crash and a missing balance top-up grant, so balance top-ups are now credited reliably.
Plan cards now show an expandable description for each task and replace "Done when" with actionable "How to validate" guidance.
Web app projects can now selectively pre-render pages as static HTML (SSG) via a simple router convention, improving load times and search engine indexing.
The source control panel now shows a Push button for newly created branches that are not yet on the remote.
After cancelling a subscription, your plan now stays fully usable until the end of the paid period.
Click the pointer icon next to the chat input or in the preview toolbar to enter Design Mode. Select any element directly in the preview to inspect and edit its properties — colors, typography, spacing, and layout — all without writing code. Changes are applied instantly via Tailwind classes and written back to your source files.
When an element is selected in Design Mode, a contextual toolbar appears below the preview with one-click tools: Fill color, Border color, Border width, Font size, Font weight, Text color, Text alignment, plus Undo/Redo and Delete. Each tool opens an inline dropdown for instant changes — no AI prompt needed.
The new color picker includes a list view with quick swatches, recent colors, and a full Tailwind palette — plus a spectrum view with an interactive HSV canvas, hue slider, HEX/RGB/HSL input modes, and EyeDropper API support. Works in both the Design Mode toolbar and the properties panel.
Below the preview, a prompt input lets you describe changes to the selected element in natural language (e.g., "make this text red" or "add a shadow"). The AI applies changes directly to the correct source file and element — no plan confirmation needed for simple edits.
The Design Mode panel shows a DOM element tree. Clicking an element in the tree highlights it in the preview and loads its properties. Selection syncs both ways — click in the preview or the tree. The element info (tag, classes, file) appears as a compact badge in the prompt bar and chat messages.
When the AI presents a build plan, you can now reject it by clicking "Nein" or typing phrases like "nein", "nicht bauen", "cancel", "skip" in both German and English. The plan is immediately discarded without calling the AI — no accidental builds.
The "Extern" tab in the media library now lets you paste or type image/video URLs directly. A live preview shows the image before you confirm. Pasted URLs are automatically accepted in picker mode.
The image property editor now has a link button next to the media library button, allowing you to type or paste any image URL directly as the source.
The publish dialog now auto-sizes to its content during the build phase instead of using a fixed 700px height.
All Design Mode section titles, toolbar labels, and property editors are now fully translated to German: Textinhalt, Bild, Filter, Layout, Abstände, Größe, Schrift, Farben, CSS-Klassen.
The "Enter to send · Shift+Enter new line" hint in the chat input now hides automatically when the chat panel is too narrow (e.g., when Design Mode is open). Based on actual container width, not viewport.
A new microphone button in the chat input and project creation screen lets you record voice messages. Your speech is transcribed via Gemini Flash and inserted into the text field for review before sending. Includes a real-time waveform visualizer and a blue shimmer animation during transcription.
After the AI modifies your project files, up to 3 contextual follow-up suggestions appear below the chat ("Continue with →"). Click to insert the prompt into the input, or Shift+Click to send directly. Suggestions are persisted and restored when you reopen the project.
A new collaborator role "Edit & Publish" lets team members edit projects and publish to production. The existing "Edit" role no longer includes publish access. Project deletion is now restricted to the project owner only.
The project list now shows when a project was last modified instead of when it was created, and sorts by most recently edited first.
Queue items can now be edited inline — click the text to open a multiline editor (Shift+Enter for new lines). A tooltip shows the full prompt text and an edit hint on hover.
Your billing name and address from Stripe are now displayed in the admin dashboard and used for invoicing. Existing customers are automatically backfilled from Stripe on the next service restart.
When you cancel your subscription (via Settings or the Stripe Customer Portal), the "Expires on" status now appears immediately instead of waiting for a webhook. Cancellations made through the Stripe portal (which uses cancel_at instead of cancel_at_period_end) are now also correctly detected.
A new "Clear chat" button lets you wipe the conversation history of any chat and start fresh. A confirmation dialog prevents accidental deletion, and the action syncs across multiple open tabs.
Token context and active task probes now run in parallel when opening a chat, noticeably reducing the time until the input field is ready. A brief grace period also prevents the loading shimmer from flickering on fast connections.
The database browser no longer exposes database credentials to the browser. All queries now run through a secure server-side proxy, improving security for projects with connected databases.
Collaborators with editor or viewer roles can now see the "Open site" button for published apps directly in the project overview.
Pressing Stop during code generation now properly saves any text already written, removes unfinished tool cards from the chat, and pauses the message queue so no queued follow-up fires after the stop.
You can now import an existing GitHub repository directly from the welcome screen and the project list. The imported repo appears as an assistant message in chat, and you can start working with AI right away.
Published web apps now have a dedicated Monitoring tab showing runtime logs and health status. Also accessible from the mobile menu.
The publish button now detects when your project has changed since the last deployment and shows a visual indicator (shimmer/split button). One click to republish with the latest changes.
When publishing, Airbuild can now generate optimized SEO metadata (title, description, Open Graph tags) using AI. The new split-preview modal lets you review and edit before going live. Published sites are also pre-rendered as static HTML for better search engine indexing.
New projects can now add a "Premium Design" chip that instructs the AI to apply extra visual polish — refined spacing, typography, animations, and a professional color palette.
The initial project questionnaire and the task plan now stream live in the chat with a shimmer animation, so you see progress immediately instead of waiting for the full response.
When the AI generates images during code creation, they now appear inline in the chat as they are produced — no need to wait until the full response is complete.
A three-checkbox consent modal is now shown at checkout, covering the right of withdrawal waiver, terms acceptance, and auto-renewal notice — in compliance with Austrian consumer protection law (FAGG).
The live task plan in the chat is now only shown after you accept it, and duplicate renderings in different chat modes have been eliminated. The plan card also auto-advances completed tasks.
Under load multiple dev-server replicas could enter `start_workspace_on_node` for the same project at the same time, producing two or three runner-side bun installs and Vite spawns that fought over the same port. The handler now acquires a Redis lock keyed on the workspace id (per-ws) AND on the project id (cross-replica) before any side effect; the loser of the race waits for the winner's status flip instead of starting its own. Same project opened in two tabs no longer creates two workspaces.
Every workspace lifecycle transition (create, sync, start, vite-up, restart, stop, delete) now emits a single-line structured log `[ws-trace event=… project=… user=… ws=… node=… elapsed_ms=… result=…]`. Across dev-server and workspace-runner the same event vocabulary is used so a single CloudWatch query can correlate a project end-to-end. Replaces a mix of unstructured `print()`/`logger.info()` lines that made post-mortem grepping hard.
After a backend redeploy the health probe could see the OLD process answering 200 for several seconds before the new process bound the port — leading to false-green status. The probe now reads the deploy version (from `/api/info`) and only returns ready when the version matches the just-deployed build sha. Same logic mirrored in workspace-runner so the dev-server / runner handshake also waits for the right binary.
When the agent fired Read/Write/Bash tools immediately after `request_integration` enabled the backend, the workspace was not yet attached and the call failed with a misleading 404. The router now waits up to 10s for the workspace to be ready, re-imports `get_workspace_for_project` post-sync to pick up the freshly persisted state, and curl is routed through the workspace-runner proxy in all environments instead of two different paths for dev vs prod.
Calling `/api/workspace/{id}/restart-vite` while Vite was already healthy used to kill it and start a fresh one — costing ~5s and a screenshot regen. The handler now probes Vite before killing: HTTP 200 within 500ms => no-op, return 200 with status "already-healthy". Reduces unnecessary HMR resets across rapid edits.
Opening a project from the dashboard previously fetched the project metadata first, then started the workspace — sequential, ~3-4s. The detail page now fires both calls in parallel and the post-restart sleep was reduced from 5s to 1s, cutting open-to-preview by ~3s in the warm-runner case.
When a publish run fails (build error, missing dep, type error), the build-error panel now offers a "Fix with AI" button that sends the error message + relevant log excerpt directly into chat as a new message — the agent picks it up and iterates a fix without the user having to copy-paste the trace.
Switching git branches now keeps the preview overlay visible until the new content has actually reloaded, so the user no longer sees the previous branch's UI during the swap. The "elapsed >= 2min"-game-overlay during long restart_frontend stays visible until the preview really comes back (5s grace), and the post-generation workspace poll now writes `frontendStatus` so the chat panel reflects ready/error correctly.
After merging a PR via the source-control modal, the workspace now checks out the base branch and re-syncs files from the merged commit — previously the working tree could stay on the now-deleted feature branch, which produced confusing "no changes" states on the next chat. Plus: chat-applied state now waits for the merge result before flipping, so a failed merge no longer marks the chat as applied.
Every label, tooltip, button, status text and error message in the Versionsverwaltung modal now ships in German and English, controlled by the existing language toggle.
The cached `node_modules.tar` blob the runner uses to skip bun-install on warm starts used to live inside the project sources directory — making git status noisy, swelling backups, and leaking into ZIP exports. It now lives under a sibling cache directory and is referenced by symlink, identically on dev-server and runner so both pick the same path.
New public, no-auth endpoint `POST /api/public/beta/apply` accepts the xaio-beta-landing form (name, email, company, role, usecase[], vision, experience, hours, referral, consent), captures IP + user-agent, sends a Brevo confirmation mail to the applicant and an admin notification to BETA_ADMIN_EMAIL (falls back to ADMIN_EMAIL). Stores all submissions for the beta-cohort selection process.
The global `projects.is_favorite` column was replaced by a `project_favorites` join table keyed by (user_id, project_id). Each collaborator now has their own favorite list on shared projects — toggling a star no longer flips it for the entire team. Migration V076 moves existing owner favorites over and drops the legacy column. The update_project endpoint accepts owner OR collaborator role for the toggle.
The cloud-local startup script previously skipped management-DB migrations, leaving developers on outdated local schemas (the "column resource_size does not exist" delete-error reproduced this). The orchestrator script now runs Flyway against `xaio_management` on every cloud-local start.
Every assistant message now has a 3-dot kebab on its action row. The popover shows two stat-pills (Clock + duration, Sparkles + AI Tasks used with the unit label muted), a "· N FILES" eyebrow, and a per-file diff list when the turn modified files. Each file row shows a TS/JS/PY/CSS language badge in neutral gray, the path, a 5-block proportional add/del bar, and +N -N counters; click the row to expand the unified diff inline. Empty popovers (no duration, no AI tasks, no files) hide the kebab entirely. Portal-rendered so it escapes the chat container overflow. Backend persists turn duration_ms and ai_tasks_used onto the iter row for survive-reload, and a new commit_sha field links each turn to its autocommit so per-file diffs are fetched via the existing /git/commits/{sha}/diff endpoint.
Messages from different days now sit beneath a centered locale-aware pill ("May 1" / "1. Mai", year only when it differs from this year). On scroll, a sticky floating pill appears at the top with the date of the message under the viewport edge; it fades 1.2s after scroll stops and hides automatically when a static separator with the same date is already visible.
Sending a new message while a turn is still streaming queues it instead of bouncing the input. The queue header sits above the input as a collapsed pill with [count] [pause] [chevron]; expand reveals the full list with a drag-handle per item, an "as next"/"als nächstes" badge on item #0, and per-item remove. When the active turn ends, the head of the queue is automatically dispatched as the next turn — unless the user paused the queue. Resume after pause also drains. State lives in the chat panel so multiple ActionBar instances on the same message stay in sync.
New "Was willst du bauen?" / "What do you want to build?" hero in Inter 500 with the last word ("bauen?" / "build?") rendered in Instrument Serif italic — the brand moment. Responsive sizes 52/40/32px with letter-spacing -1.6/-1.2/0px across breakpoints, line-height 1.05, color #111. Subtitle 15/1.5 muted #6b7280: "Ein Prompt." / "A prompt." (was "1 Prompt"). The dashed drop-zone box was removed; its function moved to a small 32×32 [+] icon button inside a single 52px footer with keyboard hint and ink-pill build button. A drag-onto-card overlay with "Drop files here" in Instrument Serif italic kicks in only during an actual drag. Keyboard hint and model picker only appear once the prompt has enough content to actually build.
The Invite-Friends modal now sends real transactional invitation emails through Brevo instead of opening a mailto. New POST /api/auth/referral/invite accepts a list of emails plus a language hint (de/en), validates + dedupes, fetches the sender's referral URL and dispatches a hand-styled HTML email per recipient: warm off-white card on a subtle gradient, +25 hero with tabular-nums, "AI Tasks" in Instrument Serif italic blue accent, ink CTA, 3-step block, footer disclaimer. Per-address result (sent / failed / invalid) is surfaced in the modal feedback line. Soft cap of 20 invitations per request. Counter switched from completed_referrals to total_referrals so the user sees real signups (label: "angemeldet" / "signed up", was misleadingly "eingeladen" / "invited").
Versionsverwaltung modal now wears the full Airbuild vocabulary: custom branch glyph instead of the GitHub icon, "verbunden" pill with green sync dot, "REPO" eyebrow with derived owner/repo slug + sync indicator, section headers per spec (chevron + icon + title + mono count-pill), "aktiv" chip in accent blue for the current branch, 28×24 white icon-buttons (Pull/Refresh/Merge/Push/PR/Checkout/Delete), 1px timeline rail with 10×10 commit nodes (accent-blue for HEAD), commit rows with mono short-hash + 13.5/500 title + author·time and a HEAD pill, file diff cards with +N / −N stats and "· Zeilen N–M" hunk headers, and a 4-col code grid (34/34/16/1fr). DiffViewer is now light-mode aware: same surface in dark, soft warm tones in light.
Media Library is now a centered 1080×720 dialog (radius 16, soft Airbuild shadow). Top bar carries a folder-icon-box, "Medienbibliothek" title in Geist 14/600 plus a mono "· N FILES · 292 KB" sublabel, a 220px search field with ⌘K hint pill, a "Von URL" ghost button, and an ink primary "Hochladen" pill. Tabs are segmented with a mono 2-digit counter; an empty tab dims to muted-soft. Asset grid uses repeat(auto-fill, minmax(180,1fr)) with 14px gap and translateY(-1px) on hover. Pills (public / private / generated) switched from saturated solids to neutral mono-outline. New footer carries an LIMITS eyebrow with the size limits and a STRG+V kbd hint. ⌘K focuses the search field; search filters the active tab live.
Team panel rebuilt as a 680px modal per spec. Header carries a Users-icon bubble, "Team" title and a mono · N MEMBERS sublabel. The inline-invite row accepts comma-separated email addresses (split + dedupe + RFC-light validate + parallel POST), reports per-address success / failure inline. Member rows show a 36px avatar (or accent-blue initial) plus name/email/"DU" tag; the Owner gets a non-clickable INHABER mono-tag with crown, everyone else gets a 28px-tall RoleSelect popover (Bearbeiten / Nur ansehen with role descriptions, a divider, and a red "Aus Team entfernen"). Removal goes through a confirm dialog with explicit pre-checks (no projectUuid, missing permission, owner-protect, self-protect) and surfaced HTTP errors — no more silent no-ops. RoleSelect popover is portal-rendered so the modal's overflow doesn't clip it.
When the model answers with an API key, password, database URL, bearer token or any other recognised secret pattern, it now renders inline as the same amber Shield-icon pill we use in user input — masked to first6...last3 with the env-var name in parentheses. Click copies the original value to the clipboard with a 1.2s "Kopiert" confirmation. Hooked into every text-bearing markdown tag (p, li, h1-h6, strong, em, blockquote, td/th, inline code, link text) plus the plain-text branch. Block <pre><code> stays untouched so multi-line snippets remain readable.
The team avatar stack in the sidebar used to filter out the current user. It now includes you — moved to the front of the list so your avatar renders rightmost in the cascading stack (closest to the Team button, top of the painting pile). Fallback initials use solid accent-blue background with a white initial instead of the previous near-transparent treatment. Bottom divider in the footer also matches the top divider's mx-2.5 inset.
User messages now carry a hover-revealed thin-icon row (Edit, Copy, Revert). Hover-only by default so the chat stays calm; full keyboard navigation supported. Combined with the kebab popover on assistant messages this gives every turn a consistent, lightweight action vocabulary on both sides.
Sidebar now lists the last 5 opened projects under the Resources section (when expanded), with rectangular thumbnails (no border / padding) and a FolderOpen fallback. The list is sourced from a client-side localStorage map (xaio:lastOpened, max 50 entries, recency-sorted) and refreshes when you switch projects. The currently open project is filtered out so the list always shows alternatives.
When the AI generates images via the generate_image tool, they now render in a small horizontal wrap-list directly inside the chat message — both in project view and dev view. Dev view places them at the bottom of the message above the timestamp, alongside the action bar.
The error / info badges on the workspace tabs (Backend, Preview, Inbox, Runs) used to overflow into the next tab's area and disappear behind its solid background. They now sit at z-20 with a 2px halo ring (white in light, ink in dark) so they're always readable, regardless of which tab is active. pointer-events: none so they don't intercept clicks meant for the tab itself.
AI Task counts now respect the document language: 1.48 in English, 1,48 in German. Token counts and detailed cost breakdowns are now hidden in prod and stage environments — they're only useful for development debugging.
`not UserReferral.referrer_bonus_granted` was being evaluated by Python at query-build time, coercing the column object to a literal True/False and silently dropping the WHERE clause. Switched to `.is_(False)` so SQLAlchemy emits the correct `WHERE referrer_bonus_granted = false`. Affects pending_referrals (used to come back as 0 or full total depending on coercion, never the actual pending set) and grant_referrer_reward (the lookup that decides whether a referrer is owed +25 AI Tasks could miss the real pending row).
When dev-server file-sync overwrites the runner's vite.config.ts (which carries the per-workspace base path), Vite serves wrong asset URLs and CloudFront returns HTML for missing /src/main.tsx. The runner's start() now re-patches vite.config and index.html every time, even when "already running". Extracted helpers _ensure_vite_config and _ensure_index_html_patched. Also: workspaces with dead nodes are now marked STOPPED instead of deleted to prevent data loss when runners are recycled.
All public marketing pages — /platform, /pricing, /about, /features, /product, /product/pipelines, /docs, /press — plus every landing component (header, footer, FAQ, how-it-works, pipelines preview, pricing preview, enterprise, social proof) now drive every visible string through the language dictionary. Switch via the existing EN/DE toggle in the header.
Top-level entry point under Platform → Übersicht. Hero with workflow strip (Describe → Clarify → Adjust → Publish), nine plain-language capability cards (project management, chat agent, live preview, database, backend & APIs, one-click publishing, versioning, design-from-screenshot, local connector), audience bands (founders / operations / marketing), cross-references to Building Blocks, Use Cases and Features, and a dark CTA.
Pricing now reflects the real plans from plan_limits.py: Free (€0 · 50 tasks), Website Builder One (€10 · 100), Website Builder Pro (€25 · 150, highlighted), Web App Builder (€50 · 200), Platform Builder (€150 · 250). Toggle now defaults to monthly, "pro Nutzer / per user" line is shown explicitly, annual shows the strikethrough monthly price. Comparison table below has 12 rows with real values.
/features dropped 10 entries with no code evidence (Magic Links, Schema Discovery, Edge Functions, Integrated DevTools, Undo/Redo, Preview Deployments, Instant Rollback, Status Page, Analytics, Secret Detection in chat-app context). Added real features: Website renewal (upload existing site, agent rebuilds modern version) and Migration from Lovable/Base44/Bolt. Renamed Monaco Editor to Source-Code Viewer to match what Airbuild actually ships.
New section "Drei Operatoren, eine These." with portrait cards for Thomas Grausgruber (CEO, formerly XOOO & Emasos), Julian Breitenecker (CMO, formerly Growth Lead · Schoolfox) and Stephan Amling (COO, formerly COO · Forterro). Mission quote attribution fixed to Thomas. Timeline shortened to four real entries: 2025 Founding, 2026 Private Beta, Q3 2026 Public Launch, Vision.
A FORCE_LIGHT_MODE flag in lib/stores/theme-context.tsx now hard-locks the marketing site to light mode regardless of system preference or stored localStorage value. The dark CSS infrastructure (variables, .dark class, dark: utilities) stays intact for a future re-enable — flip the single constant and restore the FOUC script in app/layout.tsx.
features.goLive and features.agentFlow now default to true only when NODE_ENV === "development" (next dev). Any production build defaults them to false. CI/CD pipelines need no extra wiring; an environment that does want them on can override per-build via NEXT_PUBLIC_FEATURES_GO_LIVE=true / NEXT_PUBLIC_FEATURES_AGENT_FLOW=true.
Renamed top-level "Produkt" dropdown to "Plattform", with new first item "Übersicht" → /platform. Removed Roadmap from the dropdown (gated by goLive elsewhere). Integrationen menu entry now also gated by goLive. Blog menu entry gated by goLive. Hardcoded counts ("36 Plattform-Features", "7 Pipelines", "Sechs Pläne", "Von 1.200+ Teams genutzt") removed everywhere they were stale or unverified.
New "Bewerbung" pill on /contact between Support and Presse. Both "Initiativbewerbung" CTAs on /careers now route to /contact?topic=bewerbung, which auto-selects the pill. The /api/contact subject becomes "Bewerbung-Anfrage" and the topic field routes the admin notification accordingly.
Output types (Website, Web App, Microservice, Agent Flow, Skills) are now consistently called Bausteine across the marketing site, with a separate Use Cases section under /product/pipelines for outcome-focused examples (landings, websites, web apps, agent flows, docs sites, shops, portals). The "Pipeline" term is reserved for CI/CD only.
Mobile (Flutter) and Go Microservice are no longer advertised on /product, /product/pipelines, /docs, /platform, the FAQ, or the platform tech badges. Only output types Airbuild actually ships (Website, Responsive Web App, Python Microservice, Agent Flow, Skills) are listed, and counts are computed from the visible array rather than hardcoded.
ISO 27001 and SOC 2 mentions removed from the Enterprise tile, footer system status, FAQ and /about timeline. As of this release, only GDPR (EU hosting, GDPR-compliant out-of-the-box) was claimed. ISO 27001 returned later as a started certification programme — never as a certificate we hold.
Complete rebuild of xaio.dev on a warm off-white canvas with Geist, Geist Mono, and Instrument Serif Italic typography. New hero with rotating use-cases, app-preview mock, audiences grid, seven-pipeline overview, how-it-works terminal, live prompt demo, enterprise block, pricing teaser, and FAQ accordion. All CTAs now route to /signup or /sales instead of opening modals.
The Login and Signup modals have been replaced by full /login and /signup routes in a 50/50 split layout with an editorial dark panel. Includes Google and GitHub social sign-in, email + password, live 3-bar password-strength meter, 2FA, password reset, verification-pending screen with 60-second resend cooldown. The existing backend auth endpoints and Supabase-backed session handling are unchanged.
New routes under /produkt (overview plus pipelines, agent-flow, integrationen, roadmap), the company section (/ueber-uns, /kunden, /karriere, /presse, /kontakt), the 36-feature grid at /features, the six-plan pricing page with monthly/yearly toggle and full comparison table at /preise, and a refreshed /blog feed. The top nav now offers Produkt and Unternehmen dropdowns with icon + title + description per entry.
New /sales page with a Solutions-Engineer trust row, three-step process, dark enterprise block (self-hosting, SSO/SAML, dedicated Slack, audit logs, SLA, DPA), full contact form with use-case chip toggle and interests multi-select, a 7-day calendar slot picker, sales-specific FAQ, and a final CTA. All "Mit Sales sprechen" and "Sales kontaktieren" buttons across the site now link here.
Forms on /kontakt and /sales now submit to the management /api/contact endpoint. The endpoint has a new optional topic field that routes admin notifications to the Sales, Support, Press, or Security mailbox (configurable via SALES_EMAIL, SUPPORT_EMAIL, PRESS_EMAIL, SECURITY_EMAIL — falling back to ADMIN_EMAIL). Confirmation and admin mail are now sent asynchronously via Brevo so the request no longer blocks on the roundtrip.
Flyway migration V072__add_topic_to_contacts.sql adds a topic VARCHAR(64) column plus index to the contacts table. Applied automatically on the next management deploy.
The docs frontend now uses Airbuild Blue (#2563EB), the off-white canvas, Geist + Geist Mono + Instrument Serif typography, and the shared xaio-chip, xaio-btn, and card primitives. The home page has a marketing-style hero with search pill, a 5-step journey grid, a why-Airbuild feature row, and a quick-start CTA. The top nav keeps the docs-internal links (Docs · Anleitungen · Änderungen) plus search, language, theme toggle, Login, and Start Building.
Footer and legal mentions across all public surfaces switched from "XAIO Labs GmbH" to "XAIO FlexCo".
Connect your own domain (e.g. www.my-site.com) directly to your published project. You can either attach a custom domain to an existing subdomain or publish straight to the custom domain without a subdomain. Airbuild creates a dedicated CloudFront distribution per domain, issues the SSL certificate automatically, and guides you through DNS setup. A new pending_dns status shows you exactly when your domain is ready.
Design chips now call Gemini to generate unique colors, fonts, and design tokens tailored to your specific project — no more identical-looking results. The AI also automatically suggests a matching design style based on your prompt. Four new design chips added: Glassmorphism, Luxury, Neon, and Playful. Project title preview inherits the design color and font in real time.
Create new projects directly from the sidebar without leaving your current workspace. The sidebar layout is now stable and no longer shifts when switching between views.
Projects without a backend now skip the backend build and deploy step when publishing. Publishing is noticeably faster and no longer fails on missing backend configuration.
If you start a Website project and ask for Web App features (or vice versa), Airbuild now asks before implementation if it should convert the project type — instead of silently generating the wrong stack. Conversion happens before the plan confirmation, so you stay in control.
The project cost detail view now shows how much you saved through prompt caching. Transparency on where your AI tokens actually go.
Attachments in support tickets are now clickable and downloadable via secure pre-signed URLs.
The AI agent now stops automatically when the same error keeps recurring instead of endlessly retrying. Context is deduplicated to prevent bloat, and concurrent git clones for the same project are serialized to avoid 500 errors under load.
Before starting your backend, Airbuild now validates the Python code. Syntax errors are caught immediately — no more cryptic runtime crashes during startup.
The preview console now detects and clearly reports CORS errors, so you know right away when an API call is blocked by browser security. Routine 4xx responses (e.g. validation errors) are now shown as warnings instead of errors to reduce noise.
Published sites now have a per-project Content Security Policy allowlist for external API calls. Your app can talk to the services it actually needs — without opening the door to everything.
Publishing is now rate-limited per user instead of globally. Your publishes are not blocked anymore because someone else is deploying at the same time.
Login now goes through the Airbuild backend instead of directly to the auth provider. More secure, more consistent across browsers, and easier to recover from session issues.
Fixed: clicking "Unpublish" in the confirmation dialog sometimes did nothing due to a stale event listener.
Fixed: clicking a design chip no longer re-triggers title generation. No more flash back to default when switching chips. Text after a design block is preserved when switching designs.
Users on a trialing subscription can now purchase AI Task add-ons (previously blocked incorrectly).
Upgraded axios to resolve an SSRF vulnerability, and addressed a high-severity advisory in basic-ftp.
Paste any website URL in your prompt (e.g. "Create a modern website based on example.com") and Airbuild automatically scrapes the content, downloads all images to the media library, detects the language, and builds a completely new website with the original content. Works with JavaScript-rendered sites (Wix, Squarespace, Webflow) via Jina Reader API. Sub-pages are automatically discovered and included.
After generating a migrated website, an automatic validator checks: broken image paths, missing logo in navbar, insufficient image usage (requires 50% of available), object-fit issues, and content completeness via Gemini Flash. Issues are reported in the restart_frontend result and the AI fixes them immediately.
When creating a new project (Web App Builder plan), the system automatically detects whether the prompt describes a Website or a Web App and pre-selects the correct type. Uses Gemini Flash in the same request as title generation.
Website migration images are now downloaded in parallel batches of 10 instead of sequentially. Reduces download time from 3-4 minutes to ~15 seconds for 60+ images. 90-second total timeout ensures the AI starts even if some images fail.
Website and ZIP migrations now automatically use Claude Sonnet (instead of Haiku) for stronger instruction-following. Image mapping, content completeness, and language rules are followed more reliably.
Files uploaded during project creation (logos, images) are now listed in the system prompt with usage instructions. Logos are automatically placed in the navbar. Previously, uploaded files were ignored during generation.
Portrait-only mode on mobile devices. Compact welcome screen with flexible textarea. Project type filters based on user plan. "Open Published" link in project menu. File upload error messages shown inline.
New design quality rules for generated websites: bold color contrasts, asymmetric layouts, oversized typography, proper image sizing (object-top for portraits, aspect-ratio instead of fixed heights). Premium agency feel for migration projects.
Fixed: UUID primary keys use PostgreSQL native UUID type (not String(36)). Fixed: database.py template included in copy_backend_template. Fixed: Enum types require CREATE TYPE in migrations. Fixed: no alert(), no native select/date elements. Fixed: React Router BrowserRouter wrapping.
The Publish button now shows a spinner while the API call is in progress, preventing double-clicks. Build button in welcome screen resets after errors so users can retry.
Workspaces with dead nodes are now marked as STOPPED instead of deleted. Prevents data loss when runners are recycled. Fix Errors button only shown after generation completes (not during).
The plan reconciliation tool now also adjusts AI Task balances when fixing plan mismatches. Shows adjustment details (+/- tasks) in the accounting dashboard.
Fixed published Web Apps failing with "connect() got an unexpected keyword argument sslmode/search_path". The Lambda handler now strips unsupported asyncpg connection parameters.
Upload a ZIP file of any existing project (Lovable, Manus, Base44, Bolt, Replit, or custom) and say "Migrate this". Airbuild auto-detects the tech stack, downloads images to the media library, extracts design tokens (colors, fonts, icons), and rebuilds everything in the Airbuild stack. Works with React, Vue, Svelte, Next.js, and more.
ZIP files (up to 20 MB) can now be uploaded in the chat and media library. Contents are automatically extracted and analyzed. Source code files are provided to the AI for context.
When a source project ZIP is uploaded, Airbuild auto-generates a migration reference with exact colors, fonts, icon imports, pricing data, and image paths. This ensures the AI copies content verbatim instead of inventing its own version.
The Tic-Tac-Toe game now only appears during first-time project generation. It disappears as soon as the build starts. No more flash when reopening existing projects or "Ready to build" screen after generation.
Fixed HTTP 500 errors on the Stripe webhook endpoint caused by a stripe v15 API change. Added a plan reconciliation tool to fix subscription mismatches from the outage period.
Generated apps now automatically use the browser's locale (navigator.language) for formatting dates, numbers, and currencies. A shared format.ts utility is created at project setup.
Airbuild documentation is now available at docs.xaio.dev. Accessible from the Resources section in the sidebar (all environments). Uses IBM Plex Sans/Mono fonts.
All services now expose the Git commit SHA via their /api/info or /health endpoint. Easily verify which code version is running in any environment.
The Tic-Tac-Toe game now only shows during first-time project creation and disappears when the frontend starts. No more flickering during subsequent chat changes. Variable AI difficulty makes the game more fun.
Active project type filters and view mode buttons now use Airbuild blue with a subtle border — clearly visible in both light and dark mode.
Fixed a race condition where duplicate workspaces were created due to different Redis key formats. All lookup paths now search across all key variants.
Preview iframe no longer loads when workspace is in READY/CREATED state (Vite not running yet). Prevents brief "503 not running" flash.
When AI tasks run out, the thinking spinner now stops correctly and shows a clear message with "type Continue after upgrade" hint.
Project names are now automatically generated by AI (Gemini Flash) based on your description. Type at least 3 words and the title appears after 350ms. Lock it with the lock icon or edit manually anytime.
File storage usage is now visible in the integration settings. Shows current usage vs. plan limit with a progress bar. Enforced per plan (Free: 50MB, Starter: 500MB, Pro: 5GB, Enterprise: unlimited).
The workspace status bar now shows the runner pool status for developers and admins.
When the AI service is temporarily unavailable, a friendly localized message explains the situation and tells you to continue by typing "Continue". Your work is always safe.
After an AI error, chat messages are now automatically restored from the server. No more missing messages — just refresh if needed.
Generated code now follows stricter standards: UTC timestamps everywhere, mandatory database transactions, no native browser dialogs (alert/confirm), no native select/date inputs — custom Tailwind components instead.
Uploaded profile pictures now correctly appear in chat messages instead of showing a letter avatar.
Fixed S3 upload handling in published apps: correct Lambda permissions, proper botocore packaging, and CSP headers for presigned URLs.
Duplicate any project with one click. The clone copies all source code, screenshot, and settings — but starts with a fresh chat and no git history. Enter a new name in the dialog (pre-filled with "Project Name (Clone)").
Press Escape while in the new project form to close it and return to the project list.
File checks and workspace lookups now run in parallel, saving 1-2 seconds when opening a project.
The AI now avoids redundant edits — one Edit call per file instead of repeating the same change multiple times. Also checks for existing code before copying templates on cloned projects.
Corrupt node_modules archives are now handled gracefully instead of crashing the workspace startup.
When opening a project, a task-based progress overlay shows which steps are running: Workspace, Backend, and Frontend startup — each with elapsed time, estimated duration, and a progress bar. Completed steps show a green checkmark.
"Shall we play a game?" now appears letter by letter in a typewriter effect before the Tic-Tac-Toe board reveals. The animation starts only when the preview panel becomes visible.
When code generation starts, the chat smoothly shrinks and slides left before the preview panel fades in. The Tic-Tac-Toe game appears immediately after the animation.
Upload files directly to the workspace runner for use in your project.
Project storage usage is now tracked and visible. Helps you manage resource consumption across your projects.
Students can verify their status using a university email (SWOT database) or by uploading a student ID for manual verification.
Added support for Claude Haiku 4.5, Claude Sonnet 4.6, Claude Opus 4.6, Gemini, and GPT-4o in cost tracking and billing.
Reduced the grey overlay delay from 600ms to 150ms. Preview appears almost instantly after the iframe loads.
After plan confirmation, the AI now always checks available integrations before generating code. Database and API integrations are requested upfront instead of being skipped.
Improved AI prompt for file_storage integration to prevent credential mishandling.
Fixed race conditions in database migrations using advisory locks. Schema names starting with digits (UUID-based projects) now work correctly.
Fixed OpenAPI docs routing for published apps on *.xaio.app — API docs now load correctly through the API Gateway.
Updated management and accounting services for Stripe SDK v15 compatibility. Pinned all dependencies for reproducible builds.
Prevented silent Vite startup failures with bun install retry and startup guards. Added boto3 for S3 file operations.
Every project now has an XAIO.md file that stores the design system, conventions, and important decisions. Created automatically after the first code generation. Persists across all chats — every new conversation respects the same design decisions. View it in the Files tab or Project Settings.
The AI agent can now read and visually analyze project images (JPG, PNG, GIF, WebP). Ask it to describe any image in your project.
Convert Website to Web App (and back) directly in Project Settings. Backend code is automatically initialized or archived.
Redesigned chat selector for non-technical users. Create new feature chats with a simple name input. Git branch details hidden under "Advanced". Each chat shows user avatar and message count instead of branch names.
New "Apply to project" button in the chat selector merges feature chat changes into the main project with one click. Applied chats are shown separately with a green checkmark. Status persists across page reloads.
Branch actions (Pull, Push, Sync, PR, Checkout, Delete) are now always-visible icon buttons with instant tooltips instead of a hidden 3-dot menu. PR button moved from toolbar to each branch row.
The AI no longer generates GPU-heavy infinite animations combined with blur effects. One-shot animations and gradients are used instead for cinematic designs.
All documentation pages updated. New journey-based navigation (Describe > Build > Integrate > Ship > Manage). SEO/GEO optimization with pre-rendering, sitemap, and FAQ schema.
The AI agent can now generate custom images for your project on request. Just ask in the chat, e.g. "generate a hero image for my landing page". Unsplash remains the default for stock photos.
Convert your Website project to a Web App (and back) directly in Project Settings. Backend code is automatically initialized or archived.
The AI agent can now read and visually analyze project images. Ask it to describe or analyze any image in your project.
Manage team member roles, invite collaborators via email, remove members, and transfer project ownership.
Download any asset from the Media Library with the new download button. AI-generated images now show a "Generated" badge.
Simple requests like text or color changes are automatically handled by a faster, cheaper model. Complex multi-file features still use the most capable model. This saves significant AI Tasks.
Project creation is now prompt-first: just type your idea and go. No more modal steps needed.
Project Settings now show a screenshot preview, color-coded project type badge, owner with avatar, and ownership transfer option.
All transactional emails have been redesigned with a clean, modern look including the Airbuild logo.
Fixed blank pages on published sites (*.xaio.app). PostgreSQL schema is now pre-provisioned for faster publishing.
Drag the chat panel edge to resize it. The toolbar header adjusts automatically.
Choose between Airbuild Cloud hosting (shared) and Airbuild Private (a dedicated instance in your chosen size) when publishing your project. Both run on Airbuild infrastructure in the EU.
Toolbar buttons are now disabled during AI generation instead of hidden. No more flickering or layout shifts.
Features that require a higher plan now show upgrade badges consistently across all environments.
Fixed multiple issues with database schema creation and connection handling for published Web Apps on *.xaio.app.
Fixed issues with promo code validation and discount application at checkout.
Document attachments (PDFs, images) are now resolved correctly in all chat messages, not just the latest.
Get notified when your code generation completes, team members join, or AI Tasks run low. Supports Slack, Discord, and Telegram channels.
Browse your project's PostgreSQL database directly in the workspace. Switch between Dev and Prod environments.
Upload a profile picture in Settings. Avatars are shown in the team panel and project settings.
Improved dark mode across notification settings, PostgreSQL integration modal, and database manager.
Workspace now stays alive while the AI agent is working, even when switching browser tabs.
Enter promo codes at checkout for discounts on plans and AI Task add-ons.
Signup and password change forms now show a visual password strength indicator. Strong passwords are enforced.
New community page on the marketing website with resources and links.
The marketing website has been completely redesigned with a dark premium theme and featured projects showcase.
Improved detection of passwords and secrets in chat messages to prevent accidental exposure.
Signup modal now opens automatically when visiting with a referral link. Language preference is sent during signup.
Store messages for later sending with the new queue feature below the chat input.
Smooth expand/collapse animations using framer-motion.
Messages over 10 lines or 500 characters are now automatically truncated with Show more/less.
Press Escape to close any modal or dialog across the application.
Chat input now uses the Airbuild brand color #2663eb when focused.
Fixed issues with element selection not correctly mapping to source files.
Improved preview loading times by 40% with better caching.