AI Incident Response Copilot

A portfolio-ready, multi-agent system that investigates simulated cloud incidents and keeps humans in control of remediation.

Why this project

The copilot demonstrates the operating patterns required for production AI in cloud operations: specialist-agent orchestration, tool boundaries, role-based access, approval gates, and tamper-evident accountability. Its remediations are deliberately simulated, making the full workflow safe to explore.

Architecture

flowchart LR
  U["Analyst / Commander"] -->|JWT| API["HTTP API"]
  API --> T["Triage agent"]
  T --> L["Log-analysis agent"]
  T --> M["Metrics agent"]
  T --> R["Runbook agent"]
  L & M & R --> E["Evidence + recommendation"]
  E --> H{"Human approval"}
  H -->|Approved| X["Simulated remediation tool"]
  H -->|Rejected| N["No action"]
  API -.-> A[("Hash-chained audit log")]
  MCP["MCP client"] --> TOOLS["Incident MCP server"]
  TOOLS --> T

Included scenarios

Incident	Signals	Recommended remediation
High CPU	CPU 97%, heap pressure, queue backlog	Scale checkout deployment
Failed deployment	CrashLoopBackOff, missing secret	Roll back release
Database timeout	Pool saturation, slow queries	Restart orders deployment
API latency	High p95, upstream retries	Open circuit breaker

Quick start

Requires Node.js 20 or later; there are no third-party runtime dependencies.

npm start

Open http://localhost:3000. The dashboard uses these intentionally public demo identities:

• analyst / analyst-demo — investigate incidents
• commander / commander-demo — investigate and approve/reject remediation
• auditor / auditor-demo — inspect audit data via the API

For anything beyond local demonstration, set a strong JWT_SECRET and connect a real OAuth2/OIDC identity provider. Demo passwords are not production authentication.

API

TOKEN=$(curl -s http://localhost:3000/api/auth/token \
  -H 'content-type: application/json' \
  -d '{"username":"analyst","password":"analyst-demo"}' |
  node -pe 'JSON.parse(fs.readFileSync(0)).access_token')

curl http://localhost:3000/api/incidents -H "authorization: Bearer $TOKEN"
curl http://localhost:3000/api/investigations \
  -H "authorization: Bearer $TOKEN" -H 'content-type: application/json' \
  -d '{"incidentId":"INC-1001"}'

MCP server

The newline-delimited JSON-RPC server exposes list_incidents, investigate_incident, and list_approvals:

npm run mcp

Example MCP client configuration:

{
  "mcpServers": {
    "incident-copilot": {
      "command": "node",
      "args": ["/absolute/path/to/ai-incident-response-copilot/src/mcp-server.js"]
    }
  }
}

Read-only MCP tools intentionally cannot approve or execute remediation. Approval remains behind the commander-only HTTP boundary.

Security model

• HMAC-SHA256 JWT verification with expiration checks
• RBAC roles: analyst, commander, and auditor
• Explicit pending approval object for every proposed remediation
• Commander-only approve/reject endpoints
• No direct remediation method exposed over MCP
• Hash-chained JSONL audit events for authentication, agents, tools, approvals, and execution
• UI output escaping and request size limits

This is a demonstration architecture, not a production control plane. Production hardening should add an external OIDC provider, asymmetric JWT verification/JWKS, durable transactional storage, secret management, rate limiting, policy-as-code, and real cloud APIs with least-privilege service identities.

Development

npm test
npm run check
docker compose up --build

License

MIT