HAI-MCP is the MCP implementation of Human Agent Interface.
HAI-MCP is the open-source, model-agnostic Model Context Protocol control-plane implementation of Human Agent Interface (HAI), created by Samuel Fleig.
Definition
HAI-MCP is a control plane for bounded agentic work. It exposes HAI workflow controls through MCP so compatible clients can bind agent work to mission contracts, owner gates, drift checks, and evidence-based completion.
HAI-MCP does not call a language model itself. The model remains in the client. The server manages state, rights, contracts, checkpoints, and proof requirements around the work.
Relationship to HAI
Human Agent Interface is the method: keep agentic AI work observable, bounded, verifiable, and human-owned.
HAI-MCP is one implementation of that method: a Python MCP server that gives clients a shared control surface for the HAI workflow.
What it controls
- Mission contracts: a declared problem, artifact, time limit, and done criteria.
- Owner gates: explicit human approval before selected state changes.
- Drift checks: comparison between planned or observed activity and the mission contract.
- Evidence-based completion: closing work only against verifiable criteria.
- Next-step continuity: canonical next actions across sessions and compatible clients.