Skip to content

Settings reference

This page lists every row in Settings, with the label and description the app shows. Open Settings with ⌘, (Ctrl+,) and use the search box at the top to jump to a row. Pages under Admin change how Atelier behaves for everyone on this backend; on the desktop app running on your own computer you are the admin. Pages and rows the signed-in person may not see are hidden in the app.

Setting What it does
Theme Light, Dark, or System (follow the operating system).
Chat font Atelier Serif, Atelier Sans, System, or OpenDyslexic.
Transcript text size Size of the conversation transcript text.
Transcript width Maximum width of the transcript and composer columns.
Motion Reduce animation in streaming responses and other interface elements.
Default model New chats start with this model. “Server default” follows the one set for everyone.
Start with the model I used last The model you pick in a conversation is the one the next new conversation starts with. Turn it off to always start with the one chosen above.
Default image model (only where it works) What makes your images, unless you pick another in a conversation or name one when you ask.
Default video model (only where it works) What makes your videos, unless you name another when you ask. A model that makes video from a picture only needs one to start from.
Default speech model (only where it works) What reads text aloud for you, unless you name another when you ask. Each model has its own voices.
Default music model (only where it works) What makes your songs and background music, unless you name another when you ask.
Default sound effects model (only where it works) What makes your sound effects and ambience, unless you name another when you ask.
Dictation (admins only, only where it works) The microphone in the message box: speech turns into text as people speak, with a speech-to-text service of the providers set up here.
Speech-to-text service (admins only, only where it works) Automatic uses Atelier’s own speech model. A provider can take its place: Mistral’s Voxtral Realtime (live), a local Whisper server, OpenAI, Groq, ElevenLabs or Deepgram.
Atelier’s speech model (admins only, only where it works) Nemotron 3.5 ASR from NVIDIA, run by Atelier itself: live words in about 30 languages, with nothing to set up. It is downloaded the first time someone dictates (about 715 MB) and kept.
Dictation language (only where it works) The language you speak. Auto lets the service hear it; naming it helps with short phrases and names.
Hold to record (only where it works) Holding the microphone button or the dictate shortcut records until you let go. A click still starts and stops dictating.
Microphone (only where it works) Which microphone dictation uses on this computer.
Notify me when a task finishes A system notification when a chat’s task ends while you’re in another window or app.
Notify me when a task needs me A system notification when a chat’s task waits for your answer or permission.
Setting What it does
Enable computer use (only where it works) Let the assistant in chats and Code see and use the apps you allow on this computer.
Unhide apps when Atelier finishes (only where it works) Apps hidden while Atelier had full control of your screen come back when it stops.
Move Atelier to another monitor (only where it works) With more than one monitor, Atelier moves to one the assistant isn’t working on while it uses your mouse and keyboard, so you see its progress and questions there. It comes back when the reply ends.
How Atelier uses the apps you allow (only where it works) Background: works in one app at a time in the background, without your mouse and keyboard, while you keep working. Apps you name yourself are allowed without asking; others ask first. Lists to choose from, right clicks, dragging, or an app that ignores background input need full control, which Atelier asks you for. On Linux this is “Ask per app”: you allow apps one by one and Atelier asks before it takes over your mouse and keyboard. Full control: takes over your computer as it is, any app but your denied ones, without asking.
Denied apps (only where it works) Atelier never uses these apps: asking for them is turned down without a question.
Apps allowed for good (only where it works) Apps you allowed with “Always allow on this computer”; every session may use them.
Accessibility (only where it works) macOS lets Atelier use the mouse and keyboard, and read what is under them, with this permission.
Atelier for Chrome (only where it works) The browser extension through which Atelier works in web pages, in tabs of its own in an “Atelier” group, in the background: it reads pages instead of looking at them, and never uses your own tabs. Works in Chrome, Brave, Edge and Arc. Chrome shows “started debugging this browser” while it acts in a page.
Screen recording (only where it works) macOS lets Atelier take screenshots of the apps you allow with this permission. macOS asks again from time to time.
Setting What it does
Where Atelier runs (only where it works) On this computer, or on an Atelier server you sign in to. Switching deletes nothing: switch back any time.
Share on your network (only where it works) Other devices on your network open Atelier on this computer in a browser or the Atelier app, and sign in as you with a password. They work in your chats, with your models and in your folders.
Let other devices use Atelier here (only where it works) Asks for a password and a port, then other devices on your network can sign in. Off signs them out.
Setting What it does
App shortcuts New chat, search, Settings, the sidebar and the screen lock. Change the keys or turn one off.
Message box shortcuts Upload or add files and change effort. Send, new line, stop, / and @ always work the same.
Setting What it does
Profile photo PNG, JPEG, GIF or WebP, up to 5 MB.
Full name Shown in the sidebar and to admins.
What should Atelier call you? Used in the greeting and shared with the model. Your full name is used when this is empty.
Instructions Preferences the model should keep in mind. They are added to the system prompt of every new chat.
Your ATELIER.md Your own conventions and preferences for Code sessions and chats working in a folder, in Markdown. Every new one and its sub-agents read them; where a project’s ATELIER.md disagrees, yours win.
Remote control Access your server sessions from other browsers and devices signed in to this account. Turning this off keeps sessions on their originating device, even when that session allows it.
Personality Atelier, Technical, Relaxed or your own custom mode. Style only, not model identity. Applies to new conversations; response style stays independent.
Response style How replies are written unless a chat asks otherwise.
Username (only where it works) What you sign in with.
Password (only where it works) Changing it signs you out on your other devices.
Log out (only where it works) Ends your sign-in here.
Log out of all devices (only where it works) Ends every sign-in to your account, including this one.
Content safeguards (only where it works) Extra limits on what models write for you. Your administrator manages them.
Active sessions (only where it works) Browsers and apps signed in to your account.
Devices (only where it works) Computers running the Atelier app for your account. They lend their files and tools to your sessions.
Pair mobile app (only where it works) The QR icon in the Profile heading shows a code that the Atelier app on your phone scans to sign in as you, with no address or password to type. The code works once, for five minutes, and a confirmation tells you which phone signed in.
Screen lock (only where it works) Hides Atelier until you enter your PIN or password. Your data on this computer is not encrypted by it.
PIN or password (only where it works) Asked for when Atelier starts and whenever you lock it.
Lock when inactive (only where it works) Locks Atelier after this long without typing or clicking in any of its windows.
Passkeys (only where it works) Sign in to this site with your device’s screen lock or a security key instead of your password.
Connected sign-ins (only where it works) Other sign-ins for this account in a browser, such as Authentik or Google.
Setting What it does
Keep conversations Conversations untouched for longer than this are deleted once a day. Pinned ones stay.
Temporary chats Temporary chats never appear in your history and are deleted when you leave them. Starting fresh also keeps your memory, personal instructions and projects away from the model.
Export your data A zip with all your conversations (as JSON and Markdown), projects, artifacts, files and memory.
Delete all conversations Every chat and Code session, with its messages, files and artifacts. Projects stay. This can’t be undone.
Setting What it does
Search and reference chats Let Atelier search for relevant details in your past chats.
Generate memory from chats Let Atelier remember details from your chats.
Include sensitive topics in memory Let Atelier remember details about sensitive topics like health or religious beliefs.
Import memory from other AI providers Bring what another assistant remembers about you into Atelier.
Your memory What Atelier remembers about you, your topics, areas, people and projects. Open an entry to change or delete it.
Setting What it does
Your usage Requests, tokens, images and cost of your chats and Code sessions.
Tokens by day A chart of tokens per day, split by Chat and Code.
By model Tokens, cost and average output speed of each model.
Images (only where it works) Images each model made and what they cost. A cost marked “estimate” comes from a price list, because the service didn’t report one.
Your limits (only where it works) What an admin lets you use, and how much of it you have used. Tokens, cost and requests count across all of your chats and Code sessions, including deleted ones. A cost limit counts only requests to models with a price.
Limits by person (admins only, only where it works) Everyone’s usage against their limits, with who is near or at one. Edit a person’s limits here or under Admin › Limits.
Videos (only where it works) Videos each model made, how many seconds of video that is and what it cost. A cost marked “estimate” comes from a price list, because the service didn’t report one.
Sounds (only where it works) Speech, music and sound effects each model made, how long they are in all, the characters read out and what it cost. A cost marked “estimate” comes from a price list, because the service didn’t report one.
Prices (admins only) Dollars per million tokens, per model, in place of the provider’s list price (Anthropic, OpenAI and Mistral models without one here use it). A price counts for requests made after it is set.
Setting What it does
Always-allowed actions What you approved with “Always allow” in a permission prompt. These rules apply in your Code sessions and chats, on top of the rules your admin set.
Explain each change Before each edit, Atelier says why in one short sentence, and a permission prompt shows it under the title. Every edit uses a few more tokens, whether a prompt shows or not. New sessions pick this up; a running one keeps what it started with.
Code appearance Syntax highlighting for code: one theme for the light app theme, one for the dark.
Default transcript view How Code sessions open: tool calls as usual, with thinking expanded, or with every call’s details expanded. Each session can switch in its header.
Permanent goal mode Off by default. In Code and Chat tasks, remind the assistant about unfinished todos and automatically continue a finished reply while its task list makes progress, up to four times. This can use more tokens; stops when it asks you a question.
Suggested replies After a reply in Code or a chat’s task, a short guess at your next message shows in the message box; Tab puts it in, typing sets it aside. The small model (or the conversation’s) writes it from your recent messages and the reply.
Code font A monospace font installed on your computer, for code and diffs.
Commit author name Used for commits you make from the Code view. Empty uses the repository’s git config.
Commit author email Used for commits you make from the Code view. Empty uses the repository’s git config.
Interface font Font for menus, the sidebar and panels.
Setting What it does
Folders Folders you opened for Code sessions. Their tools run on the computer the folder is on, and can run any command there as you.
SSH connections Remote machines you work on over SSH: their host, port and key, and whether Atelier keeps a password for them.
Setting What it does
GitHub Codespaces run on your GitHub account. Atelier keeps its token encrypted and uses it only for Codespaces.
GitHub OAuth app (admins only) The client id of an OAuth app with device flow enabled, so people can sign in with GitHub. Without one, they connect with a personal access token.
Codespaces Cloud machines on GitHub for Code sessions. Each one stops after its idle timeout and runs at most 12 hours at a time.
Setting What it does
Found on this computer (only where it works) Model servers answering on their default ports here (Ollama, llama.cpp, LM Studio).
Models on your computers Model servers and ChatGPT Chat on your own computers. Only you see their models, and only while that computer is connected.
Add model server Ollama, llama.cpp, LM Studio or another OpenAI-compatible server on this computer or its network.
ChatGPT Chat on this computer (only where it works) Your own ChatGPT, through the ChatGPT window of the desktop app here. Experimental.
Setting What it does
Skills Instructions and scripts that sessions load when a task needs them, or that you start with /name. They come from the folder a session works in, from the skills installed for everyone, and from plugins.
Install from GitHub (admins only) A repository or a folder in one: owner/repo, owner/repo@ref/path or a github.com URL. Skills installed with other tools, like npx skills, show up here by themselves; remove those with the tool that installed them.

Desktop app only.

Setting What it does
Version The version of the desktop app you are running.
Updates Where an update stands, with Check now, which looks for a new version right away. See Updates.
Check for updates automatically Atelier asks for the newest version when it starts and every 5 minutes. Off: it only looks when you press Check now.
Update channel Beta gets new versions before they’re released as stable.
Launch at login Open Atelier when you sign in to this computer.
Folders for sessions without one (only where it works) The server can make a new, empty folder on this computer for each session you start without a folder.
SSH hosts reached from here (only where it works) Hosts the server may connect to from this computer, with its ssh, keys and known hosts. Choosing a host here adds it.
Connectors that run a program here (only where it works) Your connectors the server may start on this computer. Adding or importing one here adds it; a changed program asks again.
Addresses reached from here (only where it works) Model servers and connectors on this computer or its network that the server may reach through it.
ChatGPT Chat on this computer (only where it works) Your ChatGPT Chat models answer through the ChatGPT window of this app.
Your instructions and skills from this computer (only where it works) Sessions in folders on this computer use your ~/.claude/CLAUDE.md, ~/.atelier/ATELIER.md, skills, agents, commands and Claude Code settings here.
Copy diagnostics Puts a plain-text summary on the clipboard to paste into a bug report: version, system, the last lines of the app log and the last error. Atelier sends nothing by itself, and the text leaves out keys, addresses, message text, file paths and names.
Data folder Where Atelier keeps its data on this computer.
Setting What it does
Report an issue Sends it to the Atelier developer with diagnostics, and a conversation and pictures if you choose. Never keys or passwords.
Request a feature An idea or a wish, for the Atelier developer’s list of requests.
Allow reports to the Atelier developer (admins only) Lets everyone here send issue reports and feature requests. Off, the report buttons say an admin turned them off.
Documentation docs.atelierhq.dev: guides, settings and what every feature does.
Copy diagnostics Versions, system and the end of the log as plain text on your clipboard, for a report somewhere else. Nothing is sent.

Admins only.

Setting What it does
Detected on this machine Local servers on their default ports and API keys in the environment.
Add provider Connect Ollama, llama.cpp, LM Studio, Anthropic, OpenAI, Mistral, GitHub Copilot, OpenRouter, a ChatGPT or Mistral Vibe plan or any OpenAI-compatible endpoint.

Admins only.

Setting What it does
Found on this computer Image and sound servers answering on their default ports here.
Image, video and audio providers Where images, videos and sounds are made, separate from the providers of text models.
Add image, video or audio provider Connect a local server such as ComfyUI, Forge or Kokoro, a cloud service such as OpenAI, Google Gemini or ElevenLabs, or a service with many models such as OpenRouter or fal.ai. The video and audio ones are listed with their own names.
Image, video and audio models Hide models, mark favorites, say what each is good at and set its steps, guidance, sampler and scheduler, for video its usual length, frame rate and resolution, or for sound its usual voice and length. The text model reads the description when it chooses.
Default image model for everyone What makes images unless a person or a conversation chooses another.
Default video model for everyone (only where it works) What makes videos unless a person chooses another or names one when they ask.
Default speech model for everyone (only where it works) What reads text aloud unless a person chooses another or names one when they ask.
Default music model for everyone (only where it works) What makes music unless a person chooses another or names one when they ask.
Default sound effects model for everyone (only where it works) What makes sound effects unless a person chooses another or names one when they ask.
Let models make images Off: conversations don’t get the tool, whichever providers are set up.
Images per turn How many images one reply can make, so text on a web page can’t run up a bill.
Images per person per day Counted across all of a person’s conversations, including deleted ones.
Spend per person per day Counts images with a known or estimated cost. Local servers cost nothing here.
Let models make videos (only where it works) Off: conversations don’t get the video tool, whichever providers are set up. Images are not affected.
Videos per turn (only where it works) How many videos one reply can make. A video takes minutes and costs far more than an image, so the default is 2.
Videos per person per day (only where it works) Counted across all of a person’s conversations, including deleted ones.
Video spend per person per day (only where it works) Counts videos with a known or estimated cost. Local servers cost nothing here.
Let models make sounds (only where it works) Off: conversations don’t get the sound tool, whichever providers are set up. Images and videos are not affected.
Sounds per turn (only where it works) How many sounds one reply can make: speech, music and effects together. The default is 6, enough for a short dialogue made line by line.
Sounds per person per day (only where it works) Counted across all of a person’s conversations, including deleted ones.
Sound spend per person per day (only where it works) Counts sounds with a known or estimated cost. Local servers cost nothing here.

Admins only.

Setting What it does
Default model for everyone New chats and Code sessions start with this model unless people choose their own.
Small model Writes titles, summaries and memory. “Same as the chat” uses each chat’s own model.
Available models Rename, hide or favorite models, and override what was detected about them.
opus Which of your models answers when an agent file asks for opus. See Models.
sonnet Which of your models answers when an agent file asks for sonnet.
haiku Which of your models answers when an agent file asks for haiku.
fable Which of your models answers when an agent file asks for fable.

Admins only.

Setting What it does
Web search Let chats look things up on the web. Nothing to set up: Atelier searches by itself.
Search engine Built-in search needs no account or key. Use your own SearXNG, Brave Search or Tavily to keep searches on a service you choose.
Test search Searches are sent to the engine that answers, as a browser would send them; news searches also go to Bing News and Google News, and programming searches to Stack Overflow, Hacker News and GitHub. The server or this computer then opens the first result pages, and your own SearXNG keeps searches off third parties.
SearXNG URL Base URL of your instance. JSON output must be enabled in its settings.yml.
Search API key Stored encrypted in the secret store, never in settings files.
Web fetch Let chats read web pages and documents by URL.
Instructions for every conversation What models are told in every chat and Code session, for every account, ahead of anything a person or project asks. Applies to new conversations. Code and chats working in a folder also follow the system AGENTS.md (Admin › Code).
Let people edit memory by hand Adds Edit to every memory entry in Settings › Memory, for changing its title, summary and details directly. Without it, memory changes only through Atelier.
Extra safeguards for everyone Off by default: models then follow only their own and their provider’s rules. When on, the categories below apply to every conversation, except for people Users sets otherwise.
Explicit sexual content Erotica and graphic sexual material. Sex education and health questions are still answered.
Hate and violent extremism Content that demeans people for who they are, and extremist propaganda. Every political view can still be discussed and argued.
Harassment of private people Threats and insults aimed at real private people, and help finding where they live or work.
Self-harm No methods or doses for someone who seems at risk; the model answers with care and points to help.
Younger people For accounts that may belong to someone under 18: no romantic role-play, and no instructions for alcohol, drugs or risky stunts.
Crisis resource A helpline or service the self-harm safeguard can point to, such as a national number.

Admins only.

Setting What it does
Default permission mode The mode new Code sessions and chats start in (a chat that would start in Plan asks instead). People can switch it in each conversation.
Allow rules Tool calls that match run without asking.
Ask rules Tool calls that match always ask first, in every mode.
Deny rules Tool calls that match are blocked, in every mode.
Additional directories Folders sessions may use besides their workspace. Reaching anywhere else asks first.
Ultracode The last effort level in Code, on every model except ChatGPT Chat, which brings its own stops: the model plans, has up to 24 sub-agents per request investigate and check its work, and uses far more tokens. Off: sessions that chose it run at Extra.
Claude Code compatibility Also read what projects and your home folder keep for Claude Code: CLAUDE.md and rules, .claude/settings.json permission rules and hooks, and the skills, commands, agents and plugins in .claude folders.
Hooks Commands that run at points of every Code session and chat task on this computer, in Claude Code’s settings format. Projects bring their own in .atelier or .claude settings; people’s own devices run only those.
Keep checkpoints for How long file versions are kept for rewinding Code sessions.
System AGENTS.md Instructions every Code session and chat working in a folder follows, for every account. Projects add their own CLAUDE.md and AGENTS.md.

Admins only.

Setting What it does
Connectors MCP servers that give sessions tools from other services. People turn them on in each session, and their tools ask before they run unless you allow them here.

Admins only.

Setting What it does
Installed plugins Bundles of skills, commands and agents. Turning one off takes what it brings out of every session. Plugins installed with Claude Code can only be turned on or off here.
Add from a marketplace A GitHub repository with a .claude-plugin/marketplace.json, as Claude Code uses: owner/repo or a github.com URL.

Admins only.

Setting What it does
Keys and secrets API keys and tokens, kept encrypted. Settings refer to them as ${secret:id}; their values never leave the backend.

Admins only.

Setting What it does
Users Everyone who can sign in, with their role and status.
Waiting for approval People who signed up themselves. Approve them to let them in.
Invite codes Each code creates one account and expires after 7 days. Only the last group of a code is shown here.

Admins only.

Setting What it does
Tokens per person per day Input, output and cache tokens of text models, counted across all of a person’s conversations, including deleted ones. A day is a UTC day and resets at 00:00 UTC. Empty means no limit.
Tokens per person per 30 days The same count over today and the 29 days before it. Empty means no limit.
Cost per person per day In dollars. Counts only requests to models that have a price (Settings › Usage › Prices); local models cost nothing here, so limit them by tokens. Empty means no limit.
Cost per person per 30 days In dollars over today and the 29 days before it, for requests to models that have a price. Empty means no limit.
Requests per person per hour Model requests in the last 60 minutes, including titles, memory updates and sub-agents. Requests under way count too. Empty means no limit.
People Each person’s usage against their limits. Raise or lower a limit for one person, lift one, or exempt them from all. Admins are not exempt unless you say so.

Admins only.

Setting What it does
Server version The version of Atelier this server runs.
Update server Installs the newest version of Atelier on this server. It restarts, and everyone reconnects by themselves.
Sign-up Invite only: admins add people or give them invite codes. Open: anyone who can reach this server can create an account, which waits for an admin to approve it.
Single sign-on Lets people sign in to this server in a browser through an OpenID Connect identity provider, such as Authentik, Keycloak or Google. The desktop app keeps signing in with passwords.

Admins only.

Setting What it does
Activity log (only where it works) Who changed what on this server, newest first: sign-ins and failed sign-ins, accounts, invites, settings, providers, secrets, connectors, plugins, models, limits and server updates. Passwords, keys and other secret values are never recorded.
Keep activity for (only where it works) Events older than this are deleted once a day. Between 7 and 365 days.
Export activity (only where it works) Saves the events that match the filters above as a JSON file, or everything recorded when none are set.

Admins only. Admins only. These are diagnostic tools.

Setting What it does
Record provider requests Lists the requests sent to model providers: address, status, timing and headers with credentials masked. Prompts and replies are never recorded.
Connector logs What each connector printed, and its connections and failures. Credentials are masked.
Dictate with Atelier’s speech model on this Mac (only where it works) This Mac dictates with its own speech recognition. On, it uses Atelier’s speech model instead: the server’s when connected to one, else this computer’s. This Mac only.
Reload connectors and extensions Reconnects every connector and reads skills, commands, agents and plugins again.