GraphOS MCP Server: graph-aware tools for your AI agent

MCP Server
, ,
1

:rocket: GraphOS MCP Server now gives your AI agent read-only access to your graph.

Point it at a graph and ask in plain language. It scans for issues across launches, schema lint, router config, and traffic metrics, then hands you a prioritized list, one prompt instead of a tour across Studio tabs.

It’s Apollo’s hosted MCP server, live at `https://mcp.apollographql.com/mcp\`. If you’ve used it for docs search or the Connectors spec, that’s the same endpoint under a new name (formerly GraphOS MCP Tools), now with graph-aware tools too.

:robot_face: What your agent can do

  • Know your graph: resolve your graph and its variants straight from your API key, no IDs to look up
  • Inspect a variant: federation version, subgraph inventory, router status and config
  • Track deployments: latest launch, launch history, and single-launch detail with composition errors
  • Watch traffic & performance: top operations plus per-operation and per-subgraph request counts, p50/p99 latency, and errors (returned as compact CSV so agents don’t drown in JSON)
  • Check quality & security: schema lint violations and persisted query list status

:bulb: Try it

  • “Scan for issues across my graph and summarize the findings by severity.”
  • “Which operations had the highest p99 latency over the last week?”
  • “Have my recent launches been stable, or were there any failures?”

It surfaces issues in what these tools can see and ranks them, rather than declaring a graph “healthy.”

:closed_lock_with_key: Scoped by your key

Pass a graph API key via the `x-api-key` header. An Observer-role key can run every tool, read-only, nothing more. Today the graph tools need a client that can send that header (Claude Code, Cursor, VS Code, Codex, Gemini CLI). OAuth for header-less clients like Claude Desktop is on the roadmap.

:link: Resources

We’d love to hear what you’re checking with it, and what you wish your agent could run next. Tell us below. :point_down: