事故调查与运维 MCP 服务器

IncidentFlow MCP

IncidentFlow MCP is an HTTP-based Model Context Protocol server for investigating and operating incidents across Kubernetes, observability platforms, and approved workspace integrations.

Hosted MCP · Documentation · Security · Contributing

What it does

• Investigates Kubernetes workloads with read-only tools.
• Reads approved Grafana dashboards and executes guarded PromQL queries.
• Brings together incident, alert, Slack, deployment, and external-status context.
• Keeps access workspace-scoped, authenticated, and auditable.

Quick start

Hosted server

The hosted server is a Streamable HTTP MCP endpoint at https://mcp.incidentflow.io/mcp. It uses OAuth 2.1 with PKCE; configure it in your MCP client and complete its browser sign-in flow.

Claude Code

claude mcp add --transport http incidentflow https://mcp.incidentflow.io/mcp
claude mcp login incidentflow

Codex CLI

codex mcp add incidentflow \
  --url https://mcp.incidentflow.io/mcp \
  --oauth-resource https://mcp.incidentflow.io/mcp
codex mcp login incidentflow --scopes mcp:read,mcp:tools:run

VS Code

Add the server to .vscode/mcp.json:

{
  "servers": {
    "incidentflow": {
      "type": "http",
      "url": "https://mcp.incidentflow.io/mcp"
    }
  }
}

See the client documentation for the sign-in command or UI. The optional VS Code installer updates this file interactively.

Local development

Prerequisites: Python 3.12 and uv.

git clone https://github.com/IncidentFlow-io/incidentflow-mcp.git
cd incidentflow-mcp
uv sync --all-groups --all-extras
cp .env.example .env
uv run incidentflow-mcp serve --reload

The server listens on http://localhost:8000 by default. Set a local bearer token with uv run incidentflow-mcp token create --name local-dev, then add the printed token to the MCP client's configuration. See authentication for the supported local and hosted flows.

To run with Docker instead:

cp .env.example .env
docker compose up --build

docker-compose.yml also includes optional local Redis and Grafana smoke-test profiles. Their operator instructions live with the deployment assets rather than in this overview.

Available tools

The server groups operational tools by task, rather than requiring a client to know every tool name up front.

All infrastructure and integration tools are read-only. knowledge_upsert is the only write-capable tool and is limited to workspace knowledge.

Inspect the exact tool inventory and JSON schemas for the installed version:

uv run incidentflow-mcp tools list --json-output

Clients can also call incidentflow_capabilities to discover the same catalog in-band.

Security model

• OAuth access tokens or platform-managed tokens authenticate every hosted request.
• Workspace identity and entitlement checks scope integration access.
• Kubernetes, Grafana, Argo CD, Slack, and external-status operations default to read-only.
• Platform allow-lists and guarded queries constrain downstream access.
• Request IDs, audit-friendly structured logs, metrics, and tracing support investigation.

Do not place tokens in query strings, commit .env files, or share local token stores. For vulnerability reporting, read SECURITY.md.

Documentation

• Authentication architecture
• Deployment and local configuration
• Observability
• Logging
• Tool response contracts
• API versioning
• Integration-status tools and metrics
• OpenAPI contract

The fern/ directory contains the internal documentation publishing setup. Run make docs-all to generate and validate the OpenAPI contract and validate its Fern configuration.

Development

make test
make lint
make openapi-validate

Before opening a pull request, follow CONTRIBUTING.md.

Repository layout

src/        MCP server implementation
tests/      Unit, integration, and contract tests
docs/       Architecture and operational documentation
openapi/    Generated HTTP API contract
docker/     Local dependency and smoke-test assets
k8s/        Kubernetes manifests and monitoring rules
scripts/    Development, validation, and client-install helpers

License

This repository is licensed under the MIT License. IncidentFlow Cloud and hosted services are proprietary.