projects
Runbook Relay
A deterministic incident-response control room where a human and an AI agent share evidence, approval state, and an audit log.
The problem
Incident dashboards are dense, stateful, and consequential. Screenshot-driven automation has to infer what a chart means, which change is selected, and whether an operator approved execution. That is too much ambiguity for an operational control surface.
What I built
Runbook Relay exposes narrow WebMCP tools for reading an incident, comparing mitigations, staging a change, executing an approved change, and resetting the simulation. The tools and the human interface use the same state transitions.
The central control is deliberate: the agent can stage a mitigation, and it cannot approve its own change. Execution fails closed until the page records explicit human approval.
What it demonstrates
- A web interface can become directly usable by an agent without disappearing from the human operator.
- Read, stage, approve, execute, and verify are separate phases.
- Tool schemas reduce ambiguity while visible receipts preserve shared evidence.
- A negative test is part of the demo: execution without approval must fail visibly.
Proof
This is a reference application, not a production operations console. A production implementation would enforce authorization and approvals server-side, bind actions to scoped identities, and persist tamper-evident audit records.