MCP
Connect AI clients to IncidentFlow operational intelligence over the Model Context Protocol.
IncidentFlow MCP is a Model Context Protocol (MCP) server that gives AI clients read‑only, workspace‑scoped access to your operational context — Kubernetes, Argo CD, Grafana/Prometheus, Slack, incidents, and workspace knowledge — behind a single authenticated endpoint.
Every response follows one stable contract, so a client written against v1
keeps working as the platform evolves.
At a glance
| Property | Value |
|---|---|
| Endpoint | https://mcp.incidentflow.io/mcp |
| Transport | HTTP Streamable (MCP), stateless |
| Methods | GET, POST (OPTIONS for CORS preflight) |
| Auth | OAuth 2.1 Bearer (recommended) or a Personal Access Token |
| API version | v1 · schema version 1.0 |
| Tools | 48 registered (44 operational + 4 meta); read‑only except one memory‑write tool |
What clients can do
- Fetch incident, integration, and workspace context.
- Inspect Kubernetes, Argo CD, and Grafana/Prometheus signals (read‑only).
- Read Slack alert threads and correlate alerts.
- Search official IncidentFlow documentation (public knowledge).
- Search workspace memory for prior incidents, RCAs, runbooks, and knowledge.
- Save operational knowledge back to workspace memory (the only write tool).
See Tools for the full catalog and API for the response contract.
Example interaction
- The client receives a new alert.
- It queries IncidentFlow MCP for integration and cluster context.
- It retrieves similar prior incidents from workspace memory.
- It searches official IncidentFlow docs for setup or runbook guidance.
- It suggests likely causes and next checks — without changing anything.
Connecting a client
IncidentFlow MCP is an OAuth 2.1 resource server. Clients that support MCP connectors (ChatGPT, Claude, and others) discover the authorization server and complete the handshake automatically:
- Add the connector URL
https://mcp.incidentflow.io/mcp. - The client reads
/.well-known/oauth-protected-resourceto locate the authorization server. - You approve access; the client receives an access token scoped to your workspace.
- The client calls tools with
Authorization: Bearer <token>.
Tokens are always workspace‑scoped — a client only ever sees data for the workspace it was authorized against.
Safety model
- Read‑only by default. 47 of 48 tools only read; the single write tool
(
knowledge_upsert) writes to workspace memory only. - No destructive actions. MCP never mutates clusters, deployments, or integrations — Kubernetes and Argo CD access is strictly observability.
- Redaction & guardrails. Logs are filtered/redacted, Grafana queries run behind server‑side PromQL guardrails, and Argo CD resource bodies and Secret data are omitted.
- Allow‑lists. Grafana dashboards and Kubernetes namespaces are limited to what the workspace has approved.
For the full safety review, see MCP Safety.
Public vs workspace tools
- Public knowledge tools search official IncidentFlow documentation and are the same for everyone.
- Workspace memory tools read and write private knowledge for the authenticated workspace only, and are never shared across workspaces.