amb

AMB is a shared message bus for AI agents working in Cursor, Codex, Claude Code, or custom workers. It gives them one shared workflow layer: threads, inbox, ACK, retry, and DLQ.

Use AMB when you want one agent to delegate work to other agents and keep the whole workflow visible in one Dashboard instead of splitting it across separate IDE chats.

This is the fastest way to try AMB in your own repository:

1. Start the published AMB stack from Docker Hub
2. Install the published MCP package in your project
3. Connect your AI client to AMB
4. Register your agents
5. Ask the orchestrator to run a workflow
6. Watch the workflow in the Dashboard

Before you start, make sure you have:

• Docker or Podman with Compose support
• Node.js 20+ in your own project
• pnpm if your project uses pnpm, or npm if your project uses npm
• an AI client that supports MCP, such as Cursor, Codex, or Claude Code
• a project that contains either .cursor/agents/registry.json or agent markdown files in .cursor/agents/*.md or .agents/*.md

For the first run, you do not need to clone the AMB repository. The quick start below uses the published AMB stack and the published @openaisdk/amb-mcp package.

Download and start the published AMB stack:

curl -O https://raw.githubusercontent.com/bizmedia/amb/main/deploy/compose/amb-compose.yml
docker compose -f amb-compose.yml up -d

amb-compose.yml already defaults to safe host ports 4333 / 4334 / 5433, so this does not collide with local AMB development on 3333 / 3334.

If those ports are also in use, override them when starting the stack:

WEB_PORT=5333 API_PORT=5334 POSTGRES_PORT=5543 docker compose -f amb-compose.yml up -d
curl http://localhost:5334/api/health

Wait until the API becomes healthy:

curl http://localhost:4334/api/health

Open:

• Dashboard: `http: