Skip to content

Releases

Atelier 0.10.4

Choose a tag to compare

@TheOriUHD TheOriUHD released this 18 Nov 14:20
  • Atelier for Chrome works for chats on a server. A desktop app connected to an Atelier server now lends the extension to it as a device service (browser, next to computer), while Computer use is on in the app. Before, only the desktop's own backend offered the Browser tool, so a chat on a server fell back to driving the Chrome app through computer use and asking for it. The server picks the computer the way it does for computer use (where the session's folder is, else the one last written from) and offers the Browser tool only while that computer's extension is connected. Older servers ignore the new service; the server needs 0.10.4.

  • No question for each step in Atelier for Chrome. The Browser tool works only in tabs of its own, so it no longer asks before opening pages, clicking or typing, in any mode. A rule you wrote (Browser) still applies, and the assistant still asks in the chat before accepting terms, sending something or paying, and never types passwords.

  • Atelier for Chrome is in the Chrome Web Store. Settings › Computer use › Atelier for Chrome opens its store page instead of describing "Load unpacked" (the folder stays available for a browser that won't take the store's extensions), the sidebar link goes there, and the docs explain Chrome's "not trusted by Enhanced Safe Browsing" note that a new publisher gets.

Atelier 0.10.3

Choose a tag to compare

@TheOriUHD TheOriUHD released this 18 Nov 12:54
  • Pairing a phone is easier to start and says when it worked. The QR button is now an icon at the right of the Profile heading at the top of Settings › Account (hover: Pair mobile app) instead of a section at the bottom, and the code is drawn black on white in every theme. When the phone has scanned it, the window turns into a confirmation such as "Philipp’s iPhone is logged in!" with "iPhone 17 Pro · Atelier 1.0" under it (new account.phonePaired event; the phone says what it is in auth.terminalPoll's device). In the Atelier app for iPhone and iPad the scanner offers a lens choice (0.5×, 1×, 2×…) on phones with more than one back camera, a flashlight button that only you turn on, a light tick when a code is read and a short run of haptic taps when you are signed in.

  • A temporary chat on the phone survives switching apps. The Atelier app for iPhone and iPad used to lose an open temporary chat when it sat in the background for more than a minute or two (the system pauses the app and drops its connection; the server then deleted the chat), and every action on it answered "That conversation no longer exists." Now the server keeps a phone's temporary chats while the app is only in the background (new app.launch), and ends them when you leave or delete them, when iOS closes the app and you open it again, or when the server restarts. If a chat is gone for any reason the app says so and opens a new one instead of showing a screen where everything fails. Needs this server release; older servers keep the one-minute rule.

Atelier 0.10.2

Choose a tag to compare

@TheOriUHD TheOriUHD released this 17 Nov 23:11
  • Models that run on your iPhone or iPad. In the Atelier app, Settings › On-device models downloads small models (LFM2.5 1.2B, Gemma 3 1B, Qwen3 1.7B and 4B) from Hugging Face onto the phone, and the model menu of any conversation then offers them under On this iPhone. The phone does the thinking; the conversation stays on your server, so you can read it anywhere. You can start a conversation on one, or take over one you started with a server model, and go back to a server model at any time. A model on a phone is only ever offered to that phone: your other devices and browsers don't list it, and a turn on it can't be started from them (they read the conversation and choose another model). It needs Atelier open on the phone while it answers, and a server of this version.

  • Keep a temporary chat. The Atelier app for iPhone and iPad has no strip across a temporary chat any more; its ⋯ menu starts with Delete chat now (deletes it and leaves) and Keep chat, which turns it into an ordinary chat. The server has a new sessions.keep call for it (the chat's own person only; its side chats come with it).

  • Sign in a phone with a QR code. Settings › Account › Sign in your phone shows a QR code, in the desktop app and in the browser. The Atelier app on a phone scans it and signs in as you, with no address or password to type. The code works once, for five minutes, and a new one ends the one before it. On the desktop it needs Share on your network; on a server it holds the address you opened Atelier at.

  • Native mobile client foundation. A SwiftUI iPhone/iPad app connects to HTTPS Atelier servers, with Keychain sign-in, browser approval, streamed conversations, projects, attachments, dictation, permission cards and native settings. Servers identify mobile sessions and deduplicate ordinary sends by client mutation ID. This is a development build, not yet a distributed mobile release.

  • Choose what the composer’s circle shows. Right-click it to select context usage or one of your provider’s reported plan limits, such as ChatGPT Plan’s five-hour or weekly window or Mistral Vibe’s monthly allowance. The choice is remembered per provider on this device; left-click still opens the details.

  • Permanent goal mode is opt-in. Settings › Code › Tasks now has Permanent goal mode, off by default. Only when enabled does Atelier remind an assistant about unfinished todos and send a finished reply back to work. It applies to Code and Chat tasks, keeps the existing progress and question checks, and allows at most four continuations.

Atelier 0.10.0

Choose a tag to compare

@TheOriUHD TheOriUHD released this 15 Nov 15:16
  • Dictation. The message box has a microphone: click it and speak, and your words appear where the cursor was as you say them, in italics until you stop; a waveform under the box follows your voice. Click again to keep the words, press Enter to send them, Esc to drop them, or just start typing. With Hold to record (on by default) holding the button records until you let go, and ⇧⌘Space (Ctrl+Shift+Space) does the same from the keyboard. The arrow beside the microphone picks the microphone, the language (Auto or one of 48, with a search) and Hold to record; the same settings are in Settings › General › Dictation.
    • Nothing to set up. On a Mac (macOS 26 or later) the Mac's own speech recognition hears you: nothing leaves the Mac. Everywhere else Atelier's own speech model does it: NVIDIA's Nemotron 3.5 ASR, live words with punctuation in about 30 languages, running on your server, or on your computer when the desktop app runs on its own. The first dictation downloads it (about 715 MB, checked against fixed checksums; the box shows the progress), then it works offline. Checked for real on macOS, Linux (also in the server's image) and Windows, in English and German.
    • Providers if you want them. Admins can choose a provider's speech-to-text instead under Settings › General › Dictation: Mistral's Voxtral Realtime (live, checked with an API key and a Vibe key), OpenAI, Groq, Together, ElevenLabs, Deepgram or a local Whisper server; or turn dictation off. Settings › Developer lets a Mac use Atelier's model (the server's) instead of its own.
    • On a Mac, macOS asks once whether Atelier may use the microphone. In a browser dictation needs https or localhost.

Atelier 0.9.3

Choose a tag to compare

@TheOriUHD TheOriUHD released this 15 Nov 01:39
  • Computer use cards for another computer are answered where you are. With Remote Control, the cards of a session on your other computer (which apps it may use, taking over the mouse and keyboard) could only be allowed in the desktop app on that computer. Now you answer them on the device you write from, with the same choices; your answer goes through the server to that computer's desktop app, which allows only the apps the card offered, at most with what it offered for each, for the session or always.

Atelier 0.9.2

Choose a tag to compare

@TheOriUHD TheOriUHD released this 15 Nov 01:13
  • Remote Control runs on the computer you control. A session in a folder on another of your computers used computer use on the computer you last wrote from: started from a laptop, it worked the laptop's screen while its commands ran on the other one. Now everything such a session does happens on the computer its folder is on, computer use included, wherever you write from. Its cards (which apps it may use, taking over the mouse) can only be allowed in the desktop app on that computer, and say so elsewhere; while computer use is off there, the session has none.

  • A new session's screen keeps its picks. Picking a model or switching thinking on one device changed the open new-session screen on your other devices at once (the model is remembered for your account). Now a screen keeps the picks it started with until you change them there; the next new screen starts from your latest picks.

  • Messages between agents join the work they belong to. "Received a message from a sub-agent", "Messaged a sub-agent" and "Background agent finished" were rows of their own that split the work into many small groups. They are now steps of the work group around them, and its summary counts them ("Read 3 files, messaged a sub-agent, 4 messages from sub-agents"); opening the group shows each one as before.

  • Connectors notice a server that comes up later. A connector that failed (an app's local MCP server that wasn't running yet) was tried again on a backoff of up to 5 minutes, so it stayed "Not connected" until someone pressed Test connection. Now it is tried again the moment its status is looked at (the message box's Connectors menu, the Connectors page) or a conversation starts, and in the background at least once a minute.

  • localhost in the desktop app means your computer, also on a server. Adding a connector at localhost or 127.0.0.1 in the desktop app connected to a server made it a server connector, which the server tried on its own machine. It now becomes your own connector, reached through this computer, and the form says so. In a browser it stays the server's, as before.

  • Switching a Code session's view keeps your place. Normal, Thinking and Verbose opened and closed every card and left you far from where you were (2,600 pixels above the end in a test). Now, at the bottom you stay at the bottom; scrolled up, the turn at the top of the window stays where it was.

  • Docs caught up with Chat and Cowork being one.

    • The prompt pages no longer describe Cowork as a mode, and the old Cowork prompt files are gone (old Cowork tasks already used Chat's).
    • /prompts/cowork/ now leads to the Chat prompt instead of a 404.
    • The usage chart counts old Cowork tasks as Chat.
  • Docs checks. The docs check now refuses a retired feature named as current (a page saying "Open Cowork…") and a redirect to a missing page. The release runs these checks before it publishes anything, not only when the docs are deployed.

  • Docs additions.

    • Three reading paths on the docs home: getting productive, building with agents, running your own setup.
    • A "What works where" page: desktop, server, browser and terminal; what needs a capable model; each OS; unofficial integrations.
    • On Providers, what can go stale, and when each provider was last checked against the live service.

Atelier 0.9.1

Choose a tag to compare

@TheOriUHD TheOriUHD released this 14 Nov 21:19
  • Reports, easier to reach and tidier.
    • The button next to your name at the bottom of the sidebar is now a bug: it opens a report with a picture of the window, like F1. Settings stays one click away in the menu under your name.
    • That menu has more: Usage, Keyboard shortcuts, Appearance, Report an issue, Request a feature, Get help, What's new, Get the apps (desktop, terminal, server, Atelier for Chrome) and Learn more (documentation, what leaves your machine, source code, license).
    • Settings has a Help & feedback page with the report buttons, the documentation, Copy diagnostics, and the admin's switch for reports.
    • An issue always carries diagnostics now; a feature request still only when you add them.
    • The Issue / Feature request switch is only as wide as its two choices.
    • A conversation is picked from a short list with a search field, instead of a menu that grew past the top of the window; a speech bubble marks chats and a terminal Code sessions. F1 or the bug button in a conversation attaches that conversation already.

Atelier 0.9.0

Choose a tag to compare

@TheOriUHD TheOriUHD released this 14 Nov 20:29
  • Report an issue or request a feature from inside Atelier. Press F1 (Settings › Shortcuts changes it), use Help › Report an Issue… / Request a Feature…, More in the sidebar, the search palette, Settings › Privacy › Reports, or /bug and /feature in the terminal app. Write a title and a sentence or two; Atelier asks for a little more when that is too short to act on and refuses padding. An issue attaches diagnostics by default; a feature request doesn't. You can also attach one conversation and up to four pictures. Each attachment can be shown before it goes:

    • Diagnostics: the computer's system, processor, memory, graphics cards with their memory and driver, free disk and monitors; the server's too, when you use one. Also settings and preferences with keys, addresses, paths and your own words left out, providers with their state, local model servers with the models they have loaded and the memory those take, connectors, computer use, and the end of the app's log.
    • A conversation: with its tool calls, results and thinking. A long one keeps how it started and its newest messages.
    • Pictures: F1 attaches a picture of the window at the moment you pressed it. Pictures are made smaller before sending.

    Keys, tokens, passwords and your account and computer names are taken out of everything. Reports go to the Atelier developer at reports.atelierhq.dev, with a proof of work and per-address limits against flooding. An admin can turn reports off (Settings › Privacy). See docs: Reporting issues.

Atelier 0.8.2

Choose a tag to compare

@TheOriUHD TheOriUHD released this 14 Nov 17:18
  • A task isn't left half done. A turn that ended while the assistant's own task list still had open items, with nothing asked of you, now goes back to work: Atelier adds a note listing what is still open and the turn continues, up to four times and only while the list keeps moving, so one that is truly stuck still ends. Some models stop mid-task to report progress whatever the prompt says (an ETS2 mod menu stopped with 5 of 10 items open after "do the whole thing on your own").

Atelier 0.8.1

Choose a tag to compare

@TheOriUHD TheOriUHD released this 14 Nov 16:16
  • A question during computer use can't be missed. When the assistant asked something after giving the mouse back, the app or game it had used stayed in front and Atelier stayed hidden behind it, so the task waited unseen. Atelier now comes forward for every question in a reply that used the computer; where Windows won't let it (a full-screen game), its taskbar button flashes and a notification says it is waiting.
  • Atelier moves to your other monitor while it works. With more than one monitor, Atelier's windows move to one the assistant isn't using while it has the mouse and keyboard, and come back when the reply ends (Settings › Computer use › Move Atelier to another monitor, on by default). Windows and apps on your other monitors are no longer minimized: only what is on the monitor it works on steps aside.
  • The assistant tests without the screen first, and doesn't ask about test licenses. It checks what it built through logs, the program's own console or debug options and scripts before reaching for the screen, sets up test servers, profiles and data itself, builds every other part when one is blocked, and accepts the license or notice of something it set up for the task (a test server's EULA, a game's SDK notice) instead of stopping to ask. Agreeing to terms for your own accounts or purchases still asks.
  • Long tasks don't stop halfway. Code no longer ends its turn just because the next step reaches outside the folder: a task you asked for that installs into or tests with something on your computer goes on, and the permission mode asks where it should. The Auto reviewer no longer blocks the agent from talking to its own sub-agents (they are reviewed like the agent) or from running an official release of a tool it downloaded for the task (a newer CMake from its GitHub releases); a message to its own sub-agent or stopping its own background work never needs review at all. When the reviewer does block something the task needs, the agent now asks you right then instead of reporting it as missing at the end.
  • Computer use finds Steam games on Windows. Games installed through Steam have no Start menu shortcut of their own, so "Euro Truck Simulator 2" read as "No app called … is on this computer". The helper now reads Steam's library folders, opens a game through Steam, and waits up to 90 seconds for its window.
  • Plan mode, answered by the reviewer. Plan mode in the mode picker has a side menu: Manual (as before) or Auto (Plan · Auto), where the Auto reviewer answers what Plan would ask. Reading anywhere goes ahead, and a command runs if the reviewer finds it only looks and changes nothing; edits wait for the approved plan as before.
  • A plan that is still being written says so. Its step reads "Writing plan…" and shows what has come in, instead of "Waiting for approval" with nothing in it; the Plan pane says the same.
  • Answers that run long stay readable. A question's own answer, a reason for a refusal and "Revise…" grow line by line up to six lines instead of scrolling sideways in one; Enter sends, Shift+Enter starts a new line.
  • A card's buttons stay on one line. Where the shortcuts are long words (Ctrl, Shift, Enter on Windows and Linux), a card shows only each button's own key unless it is wide, instead of breaking its buttons onto two staggered lines.
  • Nothing floating shows the text under it. The Comment button over a plan was see-through, as Reply once was: buttons and toolbars placed over content now always sit on an opaque surface (the copy buttons over code, the inline visual's toolbar, the memory import's Copy), and a test fails if a see-through one is placed over content again.