Apollo MCP Server

The Apollo MCP Server turns a GraphQL API into a set of tools for AI assistants. GraphQL is a query language for application programming interfaces: the client describes exactly which fields it needs and gets only those back. The server from Apollo GraphQL sits in front of such an API and exposes the operations you explicitly approve as MCP tools. MCP is the standardized interface through which assistants such as Claude Code or Cursor use external services. According to the repository the project is licensed under MIT, has around 317 stars, is not archived, and was last updated on October 23, 2026. It comes from the vendor itself, the company behind the Apollo Router and the GraphOS platform.

What can the Apollo MCP Server do?

The core is the conversion of operations into tools. According to the documentation, every named GraphQL query or mutation in your catalog becomes its own tool, with a description taken from your configuration, the operation description, a comment, or the schema. You can supply the operations as local .graphql files that reload while running, as an operation collection from GraphOS, as a manifest of stored queries, or through the persisted queries of GraphOS. Optionally you enable tools for exploring the schema: introspect looks up a type, search searches the type information, validate checks an operation without executing it, and execute runs a freely written operation. A further tool, explorer, returns a link that opens the operation in Apollo Explorer.

Installation and operation

According to the quickstart, the start goes through the Apollo command-line tool Rover, version 0.37 or later, which must be signed in with a GraphOS account, plus Node.js 18 or later for mcp-remote. With rover init --mcp you create a project, and with rover dev you start the GraphQL endpoint and the MCP server locally; the MCP endpoint is then reachable at http://127.0.0.1:8000/mcp. A Docker image can be built for operation in your own cloud. Alternatively, the repository installs the server via cargo install --locked apollo-mcp-server. The documentation names stdio for local development and Streamable HTTP for a reachable endpoint as transports. In Claude Code you connect the local endpoint with claude mcp add apollo-mcp npx mcp-remote http://127.0.0.1:8000/mcp.

Sign-in and access protection

According to the documentation, authentication works only with Streamable HTTP. The server validates signed JWT access tokens from an identity provider that supports OAuth 2.1 and, if you wish, enforces permission scopes, also per tool. Missing or invalid tokens lead to a 401 response, missing permissions to a 403. Important: the current implementation passes OAuth tokens from the clients directly to the GraphQL API behind it; the documentation points to its own security considerations for this. Custom headers such as an API key reach the API only if they are listed in forward_headers.

What is it particularly useful for?

The server is worthwhile if you already run a GraphQL API and want to give assistants a targeted slice of it. Instead of handing a model the entire API and its schema, you define a few vetted operations with meaningful names. That keeps the context small and limits what the assistant can do. Compared with a generic HTTP tool or a self-written wrapper you get ready-made building blocks for hot reload, authorization, and schema search. If you have no GraphQL API, there is nothing here for you.

Security and data flow

According to the documentation the operations are not checked against the schema before being exposed; only the endpoint behind the server validates them on execution. So review every operation before deployment. Mutations are marked as destructive by default, queries as read-only. Tools such as execute allow free operations; the mutation_mode setting controls the behavior there. What the operations return enters the AI client's context and therefore reaches the model provider.

Costs

The server itself is open source under MIT. For GraphOS, Apollo offers a free plan as well as further paid plans and enterprise tariffs; whether the MCP server is tied to a particular plan is not clearly stated on the pricing page. You will find details on Apollo's official pricing page.

Who is the Apollo MCP Server for?

It is worthwhile for developers and teams who run a GraphQL API with Apollo and want to connect assistants in a controlled way. Start with two or three read-only queries in a test environment, enable the schema tools only when needed, and turn on authentication before the endpoint becomes reachable for others.

A GraphQL API, the Rover command-line tool version 0.37 or later signed in to GraphOS (or Rust for cargo install, or Docker), Node.js 18 or later for mcp-remote, and an MCP client.

With the Apollo MCP Server you give an AI assistant targeted access to selected operations of your GraphQL API. This guide leads from the local start to the first query.

Step 1: Prerequisites

You need a GraphQL API, the Apollo command-line tool Rover version 0.37 or later (signed in with a GraphOS account), Node.js 18 or later, and an MCP client such as Claude Code or Cursor.

Step 2: Create and start a project

With rover init --mcp you create a project. Then you start the GraphQL endpoint and the MCP server with rover dev --supergraph-config supergraph.yaml --mcp .apollo/mcp.local.yaml. The MCP endpoint then runs at http://127.0.0.1:8000/mcp. You can check it with the MCP Inspector via npx @modelcontextprotocol/inspector http://127.0.0.1:8000/mcp --transport http.

npx @modelcontextprotocol/inspector http://127.0.0.1:8000/mcp --transport http

Expose operations as tools

Create one .graphql file per operation and point the configuration to the file or folder. Every named operation becomes a tool. Start with read-only queries and review each operation yourself, because the server does not validate them against the schema beforehand.

Step 3: Connect it to Claude Code

The command is claude mcp add apollo-mcp npx mcp-remote http://127.0.0.1:8000/mcp. Then check with /mcp that the tools appear.

claude mcp add apollo-mcp npx mcp-remote http://127.0.0.1:8000/mcp

Secure it

Enable the schema tools such as execute only when needed. Before the endpoint becomes reachable for others, turn on OAuth authentication with Streamable HTTP and set scopes per tool.

Locally without sign-in; with Streamable HTTP optionally OAuth 2.1 with signed JWT tokens and scopes; for GraphOS sources the variables APOLLO_KEY and APOLLO_GRAPH_REF.

Access to the approved GraphQL operations with the rights the API behind grants to the forwarded token or header; with the schema tools and execute also to free-form operations.

Results of the operations enter the AI client context and therefore reach the model provider. The server runs in your environment; OAuth tokens from clients are passed on to the GraphQL API.

Overly broad mutations or the execute tool, operations not checked against the schema in advance, passing tokens on to the API, an HTTP endpoint reachable without authentication, and prompt injection via returned data.