The pen and the mouse draw; a finger moves the page, two pinch to zoom. Move the page also with Move (V), Space held, the middle mouse button or the wheel; Ctrl + wheel zooms. The pen's buttons do what Settings → Current device says (by default: its button lassos, its eraser end rubs out). Keys: P pen, H highlighter, I line, A arrow, R rectangle, O ellipse, E eraser, L lasso, + − 0 F zoom, Ctrl+Z undo.
Projects are parts of your work (a migration, a client, the house), each with its own memory: decisions, conventions, how things are done, gotchas. Assign conversations to a project (in the chat's header) and repositories (below): the agent then gets its memory there, and adds to it as it learns.
People you know, with how to reach them. The assistant looks them up before writing, messaging or inviting someone, and adds details as you mention them.
What the assistant remembers about you between conversations. It reads these when they help and updates them as it learns; you can correct or remove anything here.
Your own instructions for the assistant. It sees each name and description, and loads the skill when a request matches.
Something happens in an integration (a pipeline fails, a mail arrives, someone writes to you in Teams…), and an agent gets a message about it and answers in its conversation. How often each integration is checked is set on its card (Integrations); only what happens after a trigger is saved fires.
Conversations besides each agent's main one: new ones you started (e.g. one per issue or merge request). Each continues where it left off. Starred ones are on top here and in the sidebar.
Agents: separate assistants, each with its own conversation, role, model and tools. Open one to chat; their settings are here.
Dev projects: a folder here or on another machine, its git repository and the account that pulls and pushes it.
Git accounts the projects pull and push with: a token for HTTPS, or an SSH key made here. One account can serve many projects.
Connect the services the assistant can work with. Each kind can be added more than once, e.g. a work and a personal account.
Everything the assistant did, or tried to do, through an integration or its own tools: what, by which agent, with what, and how it ended. Changes are kept for a year, reads for 90 days.
Everyone here uses the same assistant (its conversations, memory, reminders and projects). Admins also manage settings, integrations, git accounts, agents, projects and users.
Who agents run on, each under a name of your choice: Claude signed in to a plan (a Claude Code login, with an API key as its backup if you like), Claude with an API key, an OpenAI-compatible API (OpenAI, OpenRouter, Ollama…) or Google Gemini. Agents use the default one below unless they choose another. Besides Claude, agents get the same tools, integrations and permissions; web search and some file features are Claude Code's own.
What agents run on unless they choose otherwise (each agent's settings on the Agents page).
Checking…
Tokens and cost of every reply and routine run, counted here since this was added. Costs are estimates at API list prices; with a Pro or Max plan you pay the plan, and its limits above are what counts. The limits show for a Claude Code provider.
What can transcribe what you say and speak replies, the technical side, each under a name of your choice: Whisper or Piper here on this server (private), Google Gemini (fast and natural; it uses a Gemini provider's key, Settings → Providers), or each device's own voices. How the agent sounds is set under Voices, below.
How the agent sounds: a voice on one of the services that speak, with its speed and, on Gemini, how it speaks in your own words (tone, pace, accent). The one it speaks with is chosen under In use.
How the voice conversation waits and listens, on every device. Shorter pauses answer sooner but may cut you off mid-thought.
Reminders and replies go to every device below, always, whether or not the app is open anywhere. To get them on your phone, open the app there and enable notifications.
What agents may do with the assistant's tools by default: nothing, only look things up (read), or also change things (read & write). Each agent can differ (its settings on the Agents page). What an integration may do is set on the Integrations page.
Commands the assistant may run on this server, by their beginning (e.g. git status allows git status -s). Dangerous ones (sudo, rm -rf, git push, reading secrets…) stay blocked whatever this says. Which agents get the shell at all is set in Tools above.
E-ink: light with high contrast whatever the system's dark mode, and nothing that moves (animations, smooth scrolling, shadows), so an e-paper screen doesn't ghost. Black and white turns colours to black and greys; colour keeps them, strong enough to show on colour e-paper. Saved in this browser.
How the notepad takes this device's pen. Draw in the area above to see it. Calibrate measures your light and your firm pressure, so the lightest writing comes out thin and a firm hand reaches the thickest line; Response then bends what's between (Lighter: a light touch counts for more). Saved in this browser.
How wide the conversation and the message box are on this device; saved in this browser. Narrow screens use their whole width anyway.
On Android: makes the assistant your digital assistant, so long-pressing the power button (or the assistant gesture) opens it listening. Install this app (allow installing from your browser when asked), open it, enter this server's address and choose it under Settings → Apps → Default apps → Digital assistant app. Install the web app itself first (Install app above).
Install the CA certificate on a phone or another computer so it trusts this server’s HTTPS and can receive notifications.