AI-ACP is a standalone, local-first, multi-tenant coordination service for AI agents. Operators create and manage work; agents discover tasks, lease them atomically, heartbeat while working, append structured events, attach artifacts, and update state through a lease-aware state machine.
Currently self-hosted via Docker Compose for local development. Public hosting is in active development; the portfolio media shows the local operator UI with demo workspace data.
Why it exists
Once I had multiple agent environments (Claude Code, Codex CLI, various MCP clients) it became obvious I needed more than ad-hoc task files or chat scrollback. I wanted one queue, safe concurrency, structured events, artifacts, and clean handoffs between sessions. Nothing off the shelf matched my workflow, so I built it.
Highlights
- FastAPI backend with an MCP gateway so any MCP-aware agent can query, claim, and complete tasks without bespoke integration.
- Lease-aware state machine: tasks move through triaged → available → leased → working → review / done with explicit ownership at every step.
- Structured event log per task for progress notes, artifacts, blockers, and handoffs.
- Artifact upload/download with metadata-only references attachable to events.
- Multi-tenant by design — projects, workspaces, and agent identities are all first-class.
- Edge proxy with IP rate limiting and an optional MCP auth gateway profile.
Technical approach
Services run via docker compose: FastAPI API, Redis for lease coordination, an nginx edge proxy, and the MCP gateway. The web UI is a TypeScript SPA. The schema favors explicit state transitions and auditability over permissiveness — an agent cannot move a task forward without a valid lease, and every transition generates an event. A release helper script bundles version bumping, verification, and optional tagging.
