GraphMCP

An MCP server that speaks GraphQL. Instead of exposing one tool per endpoint, GraphMCP gives AI agents a single typed query surface: the agent reads the schema, then composes exactly the query it needs — nested relations in one call, only the fields it wants.

Ships with a mock blog dataset (users → posts → comments) so you can try it the moment you clone it.

Why

Related MCP server: anyapi-mcp-server

Quickstart

Requires Python 3.10+.

git clone https://github.com/<you>/GraphMCP.git
cd GraphMCP
python -m venv .venv

# Windows
.venv\Scripts\python.exe -m pip install -e ".[dev]"
.venv\Scripts\python.exe -m pytest -q     # verify: 12 tests

# macOS / Linux
.venv/bin/python -m pip install -e ".[dev]"
.venv/bin/python -m pytest -q

Interacting with it

Option 1 — MCP Inspector (fastest way to poke at the mock data)

npx @modelcontextprotocol/inspector .venv/Scripts/python.exe -m graphmcp.server

Opens a browser UI where you can list the tools, read the graphql://schema resource, and run queries by hand.

Option 2 — VS Code agent mode

.vscode/mcp.json is preconfigured. Open the repo in VS Code, start the graphmcp server from the MCP view, then ask Copilot agent mode things like "Who commented on Grace Hopper's posts?" and watch it discover the schema and compose queries.

Option 3 — Any MCP client (Claude Desktop, etc.)

{
  "mcpServers": {
    "graphmcp": {
      "command": "/absolute/path/to/GraphMCP/.venv/bin/python",
      "args": ["-m", "graphmcp.server"],
      "env": { "GRAPHMCP_ALLOW_MUTATIONS": "1" }
    }
  }
}

(On Windows the command is ...\GraphMCP\.venv\Scripts\python.exe.)

What the server exposes

Example: query the mock data

query($role: Role) {
  users(role: $role) {
    name
    posts(limit: 2) {
      title
      comments { author { name } text }
    }
  }
}

with variables {"role": "ADMIN"} returns, in one round trip:

{"data": {"users": [{"name": "Ada Lovelace", "posts": [{"title": "...", "comments": [...]}]}]}}

Example: write to the mock data

mutation($input: CreatePostInput!) {
  createPost(input: $input) { id published }
}

with {"input": {"authorId": "u3", "title": "Hello", "body": "..."}}. The store is in-memory — restart the server and you're back to the seed data.

Configuration

Other rails: graphql_query hard-rejects mutations, subscriptions are always rejected, and all errors come back as standard GraphQL {message, locations, path} shapes that models know how to read and repair.

How it's built

flowchart LR
    Agent["AI agent (MCP client)"] -- "MCP stdio" --> Tools

    subgraph GraphMCP["GraphMCP server"]
        direction TB
        Tools["Tools: graphql_query / graphql_validate / graphql_mutate / graphql_schema"]
        Schema["Resource: graphql://schema (SDL)"]
        Engine["Engine: parse → gate ops → validate → depth-check → execute"]
        Resolvers["Resolvers"]
    end

    Tools --> Engine --> Resolvers --> Store[("In-memory mock store<br/>(swap for DB / REST fan-out / services)")]
    Agent -. "reads schema" .-> Schema

Each layer is independently swappable:

• src/graphmcp/data.py — in-memory mock dataset. Replace with any real backend.

• src/graphmcp/schema.py — SDL with doc strings (they travel to the model) + resolver wiring.

• src/graphmcp/engine.py — execution pipeline with safety rails; no MCP dependency.

• src/graphmcp/server.py — thin MCP wiring: tools, resource, instructions.

Roadmap ideas

• Swap data.py for a real datasource (SQL, REST fan-out, microservices) — the classic GraphQL gateway pattern, now agent-facing.

• Per-field auth, query cost analysis, timeouts, result-size caps.

• Persisted-query allowlists for high-trust deployments.

• GraphQL subscriptions mapped onto MCP notifications.

Contributions and issues welcome.