A real agent, on your machine.
Not a chat that suggests code for you to copy and paste. An agent that acts in your project, under your permission rules. Everything runs local; the model comes in from the edge.
What it does, and the command for each.
Runs until it's done
Describe the task and the agent works in cycles on its own until it finishes. With a handbrake: hard caps on time and steps, and an automatic stop if something runs off the rails. "Run the tests and fix what breaks" — then go get a coffee.
A fleet working for you
Each subagent is a profile (.md) with a focus. Build your team: one for tests, one for review, one for docs. You delegate and each handles its part — in parallel, without mixing contexts.
The agents talking, live
Rooms put several agents collaborating on the same problem. They debate, divide and combine. You watch the conversation in real time and step in when it makes sense.
Connect your tools
The agent talks to external tools and integrations through MCP — the open connector standard. Add a server, search the registry, turn it on and off. Real access to what you already use.
An agent with a routine
A service wakes at the hour you set, works inside the budget you gave it, reports on your channel and goes back to sleep. The manifest says what it may do — schedule, budget, autonomy, channel — and what is a hard limit is enforced by code, never asked of the model.
The agent on your phone
Connect a bot and the agent reaches you outside the terminal: it reports when the work ends, asks when it needs a decision from you, and waits. Only the chats you allow — the allowlist starts closed, and what comes from outside it never becomes an instruction.
Everything that makes the day-to-day work.
Runs local
The process lives on your machine: session, data and context never leave it. The key stays in the OS vault — never in plain text.
Model your choice
Straight to your provider, with your own credential — API key, OAuth or a local model. You decide which one.
Edits, runs, connects
Reads and edits files, runs commands and talks to tools via MCP — in a loop, until it's solved. Every effect goes through a permission gate.
Memory between sessions
Remembers project decisions and context from one conversation to the next. See, pin or forget what it kept with /memory.
Diagnosis in one command
/doctor checks model, key, connection and tools — read-only, with a fix hint when something's off.
Themes for your terminal
Light, dark or slate. The session looks the way you like to stare at it all day.
Local where it matters, connected only where it has to be.
The session, history, project files, memory and your provider key. The agent runs locally, behind the permission gate.
The model call goes straight to the provider you configured, with your key. No middleman and no metering.