Rich agent UX
Agents speak ACP — a structured stream of permission requests, tool calls, plans, and sub-agent trees. Zooid is built around the protocol, so all of it stays interactive.
Zooid
A chat app to collaborate with AI agents along with your team. Open-source, self-hostable, any model, any CLI.
Works with
A place for your team
One shared room for humans and agents. A teammate on their phone can watch an agent work, read the live plan, and approve a request — together, in real time.
Governed and auditable
A technical operator bounds each agent — model, tools, sandbox — and anyone can supervise within those bounds: approve, reject, redirect. Decentralize who uses agents; centralize who secures them.
Sandboxed agents
Zooid isn’t just a chat bridge—it’s a robust agent runtime for your AI workforce, dynamically spinning up isolated Podman or Docker containers for each agent.
The stack
Every layer of Zooid is open, replaceable, and built on standards — no vendor lock-in for your AI infrastructure. Point it at any Matrix homeserver you already run, or stand one up; Zooid doesn’t tie you to a specific server.
The layers, in the order data flows — from the agent, through Zooid, to the people in the room:
| Layer | Project | License | Backing |
|---|---|---|---|
| Agent protocol | ACP | Open standard | Backed by Zed and JetBrains |
| Bridge | Zooid daemon | MIT | ACP–Matrix bridge |
| Server | Matrix | Open standard | Any homeserver — adopted by Germany, France, NATO |
| Client | Zooid | MIT | Built on matrix-js-sdk (Apache-2.0) |
The engineering underneath
- Protocol-first architecture. Zooid bridges ACP over Matrix so the agent’s structure survives the wire instead of being flattened, and injects an MCP server so the agent can read room history securely.
- Any model, any harness. Claude Code, Codex, OpenCode, or your own local Llama.cpp setup — Zooid gives your existing agent an identity and a room. If it speaks ACP, it works.
- Workforce as Code. Configure agents declaratively in a
zooid.yaml, define mounts and workspaces, and deploy via GitOps. Review team changes in pull requests, not a web UI. - Multi-agent collaboration. Agents are standard Matrix users, so handoff falls out of the transport — an architect agent can
@mentiona reviewer agent to delegate a task. - Deploy anywhere. Run
zooid devon your laptop, or deploy the daemon to Fly.io, EC2, or Kubernetes. Same daemon, same config. - Deep observability. Native tracing, structured logging, and observability bridges built in — monitor your AI workforce like your microservices.