Apollo MCP Server

Apollo MCP Server is the official bridge between GraphQL APIs and AI agents, from the team behind Apollo Client, Apollo Server and Apollo Federation. Rather than asking you to hand-write a tool wrapper per endpoint, it generates MCP tools from artefacts you already have.

Three ways to produce tools. Point it at individual .graphql operation files and each operation becomes a tool with a schema derived from its variables. Point it at a Persisted Query Manifest from Apollo GraphOS and your already-approved, already-reviewed operations become the agent's tool surface — which is the safest option, because the agent can only run queries that passed your normal review. Or let it introspect the schema dynamically when you want flexible discovery over a graph that is still moving.

That first-versus-second distinction is the design point worth understanding. Exposing raw introspection to an agent gives it the whole graph, including expensive queries and mutations nobody vetted; driving tools from a persisted-query manifest keeps the blast radius equal to what your team already ships. Apollo supports both and lets you choose per deployment.

Beyond tools. The server also supports MCP prompts for guided multi-step workflows, and MCP Apps for building interactive experiences on top of a graph.

Running it. rover init --mcp scaffolds the configuration, and rover dev --supergraph-config supergraph.yaml --mcp .apollo/mcp.local.yaml runs it alongside a local graph (Rover CLI v0.37+). For deployment there is a mcp.Dockerfile and a prebuilt container at ghcr.io/apollographql/apollo-mcp-server, driven by APOLLO_KEY and APOLLO_GRAPH_REF. Default transport is HTTP on port 8000, configurable via APOLLO_MCP_TRANSPORT__PORT. The server is written in Rust and can also be built from source with cargo build.