The first cut of OpenBot. Alpha, and under active development: it runs on a laptop, there are rough edges, and things will move.
What it is
Bots are teammates you hand real work to. Each one gets a computer of its own, a container holding its own Chromium with its own logins, its own filesystem, and only the tools you granted it. No Bot can read another's files or reuse another's sign-ins.
Every action a Bot takes is decided before it happens and recorded after. Each page opened, file written, MCP tool called and component drawn passes a policy you write, and lands in a trail that keeps what was allowed and what was refused, every refusal naming the rule that refused it.
In this release
- Bring any AG-UI agent, on a framework or hand written. A proof-of-concept Bot and a LangGraph Bot ship with it.
- A computer per Bot, created by the supervisor, with a persistent browser profile and workspace.
- The gateway: CEL policy with deny evaluated before allow, fail closed, and the audit row written before the action runs rather than after it succeeds.
- Take the wheel. A Bot that reaches something it should not do alone stops and asks, you drive its browser, and hand it back.
- Components instead of only prose, authored in the browser and published without a rebuild.
- Governed MCP, with catalogue entries for Atlassian, Box, Slack, Salesforce and ServiceNow, plus personal and deployment skills.
- Durable threads and memory through CopilotKit Intelligence.
- Optional Google sign in, or a single-administrator dev mode.
Running it
Docker, Bun 1.3+, a CopilotKit Intelligence project and licence, and a model key. Then:
cp .env.example .env
bun install
bash scripts/start.shThe README has the full quick start, and docs/architecture.md has the diagram and the detail behind it.
Known limits
- Local only. There is no hosted deployment yet.
OPENBOT_DEV_NO_AUTHis the default and admits every request as one administrator. Wire up Google sign in to change that.- The SPIRE services in the compose file are optional and are not started.