AgentLoopMCP
One shared loop server. Many agents. Real continuity.
AgentLoopMCP is the transactional loop-state and coordination service for recurring LLM agents. It stores each agent's private handoff state, run journal, leases/fencing, and shared coordination notes in one PostgreSQL database — so many agents share one install without process-wide identity.
Who it is for
- OpenCode / HOS scheduled agents that must remember the last successful cycle
- Multi-agent stacks that need a shared mailbox without a workflow engine
- Operators who refuse one-container-per-agent topology
Start here
- Register identity —
loop_workspace_registerthenloop_agent_register - Run a cycle — start → work → complete with fencing tokens
- Technical architecture — layers, tools, errors (login required)
What you get
| Capability | Outcome |
|---|---|
| Pair identity | Every call carries workspace_name + agent_name — no AGENT_LOOP_DEV_* |
| Registration tools | Idempotent workspace and agent registration |
| Run lifecycle | Start / heartbeat / complete / fail with fencing + idempotency keys |
| Private continuity | Last successful private_state returned on the next start |
| Shared notes | Urgent publish now or atomic notes on complete |
| Checkpoints | Per-source cursors with compare-and-swap versions |