# mcp-graphql-bridge [![npm version](https://img.shields.io/npm/v/mcp-graphql-bridge.svg)](https://www.npmjs.com/package/mcp-graphql-bridge) [![CI](https://github.com/murilojrpereira/mcp-graphql-bridge/actions/workflows/ci.yml/badge.svg)](https://github.com/murilojrpereira/mcp-graphql-bridge/actions/workflows/ci.yml) [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE) [![Node.js >= 20](https://img.shields.io/badge/node-%3E%3D20-brightgreen)](https://nodejs.org) A generic MCP (Model Context Protocol) server that bridges any GraphQL API to Claude Code. It introspects your GraphQL schema and exposes each query and mutation as an individual tool, letting Claude interact with your API directly. ## How it works On startup the server will: 1. Look for a `schema-introspection.json` file in the working directory (fast, no network call) 2. If not found, run live introspection against `GRAPHQL_INTROSPECTION_URL` 3. Register one tool per query (`query__`) and one per mutation (`mutation__`) 4. Always register a generic `execute_graphql` fallback tool and a `get_type_details` explorer tool ## Requirements - Node.js >= 20 ## Setup ### Step 1: Install #### Option A: Install from npm (recommended) ```bash npm install -g mcp-graphql-bridge ``` #### Option B: Clone and build from source ```bash git clone https://github.com/murilojrpereira/mcp-graphql-bridge.git cd mcp-graphql-bridge npm install npm run build ``` ### Step 2: Configure environment variables | Variable | Required | Description | |---|---|---| | `GRAPHQL_API_URL` | No | Endpoint used for queries and mutations. Defaults to a public demo API ([countries.trevorblades.com](https://countries.trevorblades.com/graphql)) if unset — replace with your own for real use. | | `GRAPHQL_INTROSPECTION_URL` | No | Endpoint used for schema introspection. Defaults to `GRAPHQL_API_URL` if unset. | | `GRAPHQL_TOKEN` | No | Bearer token for GraphQL authentication (used for query/mutation execution). Omit for public APIs. | | `GRAPHQL_INTROSPECTION_TOKEN` | No | Bearer token for schema introspection, if it requires different credentials than execution (e.g. a separate schema registry). Defaults to `GRAPHQL_TOKEN` if unset. | | `MCP_AUTH_TOKEN` | No | Bearer token required by the hosted `/mcp` HTTP endpoint when `MCP_TRANSPORT=http` | | `GRAPHQL_MAX_TOOLS` | No | Maximum number of query/mutation tools to register. Queries are prioritized over mutations when truncating. Default `128`. | | `GRAPHQL_INCLUDE_MUTATIONS` | No | Set to `false` to exclude every mutation field entirely, for a read-only deployment. Default `true`. | | `GRAPHQL_MAX_RETRIES` | No | Retries (0–5) for `429`/`502`/`503`/`504` responses, honoring `Retry-After` when present. Default `0` (disabled). | For schemas with hundreds of fields (GitHub's GraphQL API has 284 root fields — 32 queries, 252 mutations), `GRAPHQL_MAX_TOOLS` and `GRAPHQL_INCLUDE_MUTATIONS` are what keep registration bounded and predictable. If the cap truncates the schema, stderr logs exactly how many queries/mutations were registered vs. available. No configuration is required to try the server — with nothing set, it starts against the public demo API above and logs that it's doing so. See [`docs/architecture.md`](docs/architecture.md) for the full token model and why the GraphQL endpoint is fixed per deployment rather than a per-request parameter. You can set these in a `.env` file at the project root: ```env GRAPHQL_API_URL=https://your-api.example.com/graphql GRAPHQL_INTROSPECTION_URL=https://your-api.example.com/graphql GRAPHQL_TOKEN=your-bearer-token ``` Or pass them directly via the `claude mcp add` command (see below). ### Step 3: (Optional) Pre-generate schema snapshot By default the server introspects your schema live on startup — no file needed, and it automatically retries at a shallower query depth if your API rejects the full-depth attempt (some APIs, especially CDN-fronted ones, enforce a query depth limit). Use this step only if your API has introspection disabled entirely in production, or you want faster startup times: ```bash curl -s -X POST https://your-api.example.com/graphql \ -H "Content-Type: application/json" \ -H "Authorization: Bearer your-bearer-token" \ -d '{"query":"{ __schema { queryType { fields { name description args { name description defaultValue type { kind name ofType { kind name ofType { kind name ofType { kind name ofType { kind name ofType { kind name ofType { kind name } } } } } } } } type { kind name ofType { kind name ofType { kind name ofType { kind name ofType { kind name ofType { kind name ofType { kind name } } } } } } } } } mutationType { fields { name description args { name description defaultValue type { kind name ofType { kind name ofType { kind name ofType { kind name ofType { kind name ofType { kind name ofType { kind name } } } } } } } } type { kind name ofType { kind name ofType { kind name ofType { kind name ofType { kind name ofType { kind name ofType { kind name } } } } } } } } } } }"}' \ > schema-introspection.json ``` If your API rejects this with a depth/complexity-limit error, shrink the `ofType { ... }` nesting (each level resolves one more `NonNull`/`List` wrapper — most real-world types need 2-3 levels; only doubly-wrapped lists like `[[Int!]!]!` need more). ## Adding to Claude Code ### Option A: User scope (just for you) **If installed from npm:** ```bash claude mcp add --transport stdio \ --env GRAPHQL_API_URL=https://your-api.example.com/graphql \ --env GRAPHQL_INTROSPECTION_URL=https://your-api.example.com/graphql \ --env GRAPHQL_TOKEN=your-bearer-token \ graphql-bridge -- mcp-graphql-bridge ``` **If cloned from source:** ```bash claude mcp add --transport stdio \ --env GRAPHQL_API_URL=https://your-api.example.com/graphql \ --env GRAPHQL_INTROSPECTION_URL=https://your-api.example.com/graphql \ --env GRAPHQL_TOKEN=your-bearer-token \ graphql-bridge -- node /absolute/path/to/mcp-graphql-bridge/dist/index.js ``` > **Important:** Make sure to use `mcp-graphql-bridge/dist/index.js` (the compiled output), not `mcp-graphql-bridge/index.js`. The TypeScript source must be built first with `npm run build`, and the entry point is in the `dist/` folder. ### Option B: Project scope (shared with your team via `.mcp.json`) ```bash claude mcp add --transport stdio --scope project \ --env GRAPHQL_API_URL=https://your-api.example.com/graphql \ --env GRAPHQL_INTROSPECTION_URL=https://your-api.example.com/graphql \ --env GRAPHQL_TOKEN=your-bearer-token \ graphql-bridge -- mcp-graphql-bridge ``` > **Note:** Use absolute paths. All `--env` and `--transport` flags must come before the server name. ### Verify the connection ```bash claude mcp list ``` Then in a Claude Code session, run `/mcp` to see available servers and tools. ### MCP Registry This server is published to the official [MCP Registry](https://registry.modelcontextprotocol.io) as `io.github.murilojrpereira/mcp-graphql-bridge`, so MCP clients that browse the registry can discover and install it from there. ## Examples Two worked walkthroughs — a small public schema with no configuration needed, then a large, real enterprise-scale schema requiring auth and tool-count limits. ### Example 1: Countries API (small schema, no auth) This is the zero-config default — nothing to install or configure beyond the server itself. 1. Add the server with no environment variables at all: ```bash claude mcp add --transport stdio graphql-countries -- mcp-graphql-bridge ``` 2. Restart Claude Code (or run `/mcp` to confirm `graphql-countries` is connected). You should see tools like `query__country`, `query__countries`, and `query__continents`. 3. Ask Claude: > Using graphql-countries, find the country with code "BR", then list its continent's other countries. Claude calls `query__country({ code: "BR", __fields: "{ name continent { code name } }" })`, then `query__continent` or `query__countries({ __fields: "{ name }" })` filtered by the result. 4. Try an invalid code to see error passthrough: > Look up the country with code "ZZZ". Returns the GraphQL API's own error text — the bridge passes it through rather than masking it. ### Example 2: GitHub GraphQL API (large schema, auth + tool limits) GitHub's GraphQL API has **284 root fields** (32 queries, 252 mutations) — far more than the `GRAPHQL_MAX_TOOLS` default of 128, and it needs a token for every request, including introspection (unlike GitHub's REST API, which allows some anonymous reads). 1. Add the server, scoped to read-only access: ```bash export GH_TOKEN=ghp_your_personal_access_token # or: source a gitignored .env file first claude mcp add --transport stdio graphql-github \ --env GRAPHQL_API_URL=https://api.github.com/graphql \ --env GRAPHQL_INTROSPECTION_URL=https://api.github.com/graphql \ --env GRAPHQL_TOKEN=$GH_TOKEN \ --env GRAPHQL_INCLUDE_MUTATIONS=false \ graphql-bridge -- mcp-graphql-bridge ``` `GRAPHQL_INCLUDE_MUTATIONS=false` registers all 32 (read-only) queries and zero mutations — comfortably under the cap, and a meaningfully safer default for an AI agent than exposing all 252 write operations. 2. Ask Claude: > Using graphql-github, look up the repository facebook/react and tell me its star count. Claude calls `query__repository({ owner: "facebook", name: "react", __fields: "{ name stargazerCount }" })`. 3. To also reach mutations, drop `GRAPHQL_INCLUDE_MUTATIONS=false` and raise the cap (`GRAPHQL_MAX_TOOLS=400`), understanding that this exposes write access to your GitHub account scoped to whatever permissions your token has. ## Available tools | Tool | Description | |---|---| | `query__` | One tool per GraphQL query field | | `mutation__` | One tool per GraphQL mutation field | | `execute_graphql` | Generic fallback — run any query or mutation (mutations rejected if `GRAPHQL_INCLUDE_MUTATIONS=false`) | | `get_type_details` | Explore fields of a specific GraphQL type | All per-operation tools accept a special `__fields` argument where you can provide a custom GraphQL selection set (e.g. `{ id name status }`). If omitted, only scalar fields are returned. **Per-call auth override**: every tool (including `execute_graphql`) also accepts `bearer_token` and `custom_headers` arguments. If provided, they override `GRAPHQL_TOKEN`/no-auth for that single request only, letting Claude switch credentials per call without restarting the server. ## Security - **The target API is fixed per deployment, never a per-request parameter.** Individual tool calls can override *credentials* (`bearer_token`, `custom_headers`) but never the destination host — `GRAPHQL_API_URL` is set once at deployment time. A shared server that let callers redirect it to an arbitrary destination would be a Server-Side Request Forgery (SSRF) primitive; this design rules that out by construction. - **Configured and per-call secrets are redacted from every response** before it reaches the calling LLM. - **`GRAPHQL_INCLUDE_MUTATIONS=false`** excludes every mutation field from registration for a genuinely read-only deployment — a meaningful trust boundary GraphQL's type system already encodes, rather than relying on token scope alone. This is enforced for `execute_graphql` too: it parses the query and rejects any mutation when this flag is off, rather than only omitting the convenience `mutation__*` tools while leaving the generic fallback able to run anything. - **`MCP_AUTH_TOKEN`** gates the HTTP transport's `/mcp` endpoint for public-routable deployments; requests are capped at 10MB. See [`docs/architecture.md`](docs/architecture.md) for the full design rationale and [`SECURITY.md`](SECURITY.md) to report a vulnerability. ## Docker ### Build the image ```bash docker build -t mcp-graphql-bridge . ``` ### Add to Claude Code via Docker ```bash claude mcp add --transport stdio \ --env GRAPHQL_API_URL=https://your-api.example.com/graphql \ --env GRAPHQL_INTROSPECTION_URL=https://your-api.example.com/graphql \ --env GRAPHQL_TOKEN=your-bearer-token \ graphql-bridge -- docker run -i --rm \ -e GRAPHQL_API_URL -e GRAPHQL_INTROSPECTION_URL -e GRAPHQL_TOKEN \ mcp-graphql-bridge ``` > **Note:** The `-i` flag (no `-t`) is required — it keeps stdin open for the MCP stdio protocol. ## HTTP deployment For hosted MCP access, run the HTTP transport instead of stdio: ```bash docker build -f Dockerfile.http -t mcp-graphql-bridge-http . docker run --rm -p 8080:8080 \ -e GRAPHQL_API_URL=https://your-api.example.com/graphql \ -e GRAPHQL_INTROSPECTION_URL=https://your-api.example.com/graphql \ -e GRAPHQL_TOKEN=your-bearer-token \ mcp-graphql-bridge-http ``` Health checks are available at `/health`; MCP requests are served at `/mcp`. For public-routable deployments, set `MCP_AUTH_TOKEN` and configure clients to send `Authorization: Bearer ` to `/mcp`. See [`docs/deployment.md`](docs/deployment.md) for AWS, Cloudflare, and other container hosting options. ## Development ```bash npm run dev # watch mode: rebuilds and restarts on file changes npm run build # one-off TypeScript compile npm start # run the compiled server ``` See [`CONTRIBUTING.md`](CONTRIBUTING.md) for contribution guidelines and the maintainer release process. ## Troubleshooting ### Error: Cannot find module '.../index.js' If you see an error like: ``` Error: Cannot find module '/path/to/mcp-graphql-bridge/index.js' ``` You are pointing to the wrong file. The TypeScript source must be compiled first, and the entry point is in the `dist/` folder: **Correct path:** `/path/to/mcp-graphql-bridge/dist/index.js` **Wrong path:** `/path/to/mcp-graphql-bridge/index.js` **Fix:** 1. Ensure you ran `npm run build` (creates the `dist/` folder) 2. Update your MCP configuration to use the full path ending in `/dist/index.js` ### Schema introspection fails If the server starts but shows "Schema introspection failed", your GraphQL API may have introspection disabled in production. Use the curl command in step 3 of Setup to pre-generate a `schema-introspection.json` file. ### Tools not appearing in Claude Code 1. Run `claude mcp list` to verify the server is registered 2. Run `/mcp` in a Claude Code session to see available tools 3. Check that your GraphQL API's environment variables are set correctly (`GRAPHQL_API_URL`, `GRAPHQL_INTROSPECTION_URL`, `GRAPHQL_TOKEN`) — these are optional and default to a public demo API, so if tools still aren't appearing with your own API configured, check its credentials and endpoint URLs