Operator-facing MCP

Connect Sentinela to agents without pretending it should run itself.

Sentinela now ships with a local Model Context Protocol surface for operator copilots, queue review, escalation drafting, and ingestion monitoring. It is designed to make agents useful in the response loop without giving them unsafe autonomy over public incident workflows.

Run the Sentinela Ops MCP server

The MCP package lives inside the repository and runs over stdio, which makes it easy to connect from Codex or another MCP-capable client while keeping Sentinela's operator context close to the code and database.

Command

pnpm install
pnpm mcp:ops

Included Codex skills

  • sentinela-triage
  • sentinela-ops-health

Prepare operator-ready workflows

Expose the live queue cleanly

Agents can inspect current reports, pull full incident context, and reason over contributions without screen-scraping the product.

Prepare operator-ready workflows

Queue triage, incident review, and authority notification drafts can be orchestrated through explicit tools instead of brittle internal prompts.

Keep safety boundaries visible

The first MCP surface is intentionally read-only except for draft generation. It does not send emails or mutate incident state.

What the current MCP server exposes

The first release stays narrow on purpose. It focuses on report inspection, escalation preparation, and worker observability.

list_reports
Returns the current report queue with filters for status, incident type, and time window.
get_report
Loads one incident with images, score, weather snapshot, and contribution history.
draft_government_notification
Builds a structured government-facing draft using the same incident framing used inside the product.
get_worker_health
Returns overall worker health, Redis connectivity, scheduler state, and ingestion posture.
get_collector_status
Shows one collector's state or a compact summary of every collector behind the live map.

What this is actually for

  • Review fresh community reports and separate high-signal incidents from vague submissions.
  • Generate escalation-ready copy for a human operator before anything is sent to authorities.
  • Monitor whether ingestion feeds are healthy, degraded, stale, or failing before map trust erodes.

Useful agents need constraints.

Sentinela is part of an emergency workflow. The agent layer should assist operators, not silently take authority it has not earned.

  • No report-state mutations are exposed through the MCP server yet.
  • Notification generation stops at draft output. Sending remains a human decision.
  • Worker health is observable, but collectors are not remotely controlled from this surface.
  • The current skills are optimized for triage and ops review, not autonomous dispatch.