product

How Kojo works.

A session is durable; a run is bounded. Everything below is what the product does today; anything marked preview is built but not yet available to everyone.

On this page: sessions · the computer · runs · repositories · API · limits

Sessions and members

Members have roles. API keys and apps join as automation, never as owners. A reviewer or above answers the agent's approval request from any device; a step nobody answers in time is denied. Anyone allowed can queue the next message while the agent works, or interrupt it.

Role	Can
`viewer`	read the conversation and the activity
`reviewer`	…and approve or deny the agent's requests
`editor`	…and send, queue and interrupt; use the terminals and open any preview
`owner`	…and manage the session and its members
`automation`	read and send messages; how API keys and apps join, never higher

The computer

cpu         4 vCPU
memory      8 GB
checkout    /workspace/<repository>
sleeps      after 15 idle minutes, memory and processes kept
wakes on    a message, a preview, a keystroke

Shared terminals. tmux sessions everyone in the session sees; they outlive browser tabs.
Browsers. A browser for every run, so the agent can see what it built.
Previews. One origin per preview; signed links expire in 15 minutes; viewers reach only the session's declared preview port.

Runs and recovery

A run lasts up to two hours. The branch is pushed as it goes, the transcript is exported and resumed by the next run, and a gone computer is rebuilt from the pushed branch. A usage limit or an outage moves the run to the next model route in your chain, in the same conversation.

Autonomy modes: Ask, Auto-edit, Full auto.

Repositories and pull requests

A session holds every repository you pick; Kojo never guesses them. Its work publishes as linked draft pull requests, one per repository, opened in order and retried if one fails. A merge always waits for a person, and the agent opens pull requests through Kojo, not with a token of its own.

The API and the local client

An API with the same surface as the app: scoped API keys, a catalog endpoint listing each operation with its JSON Schema and the scope it needs, read-only MCP tools, and a local client that lets your own Claude Code join a cloud session. An agent can start work for other agents up to 3 levels deep and 5 at once, by default.

What it isn't, yet

Claude Code only. Open models plug in through Claude Code; Codex and Cursor do not.
One active run per session. Parallel work means more sessions, or a fork of one.
A web app. It works on desktop and phone; there is no native app and no local worktree mode. For one person on one laptop, a local tool is still faster.
Hosted only. No self-hosting and not open source.

Start a session