GraphQL gateway for LLMs. Handles huge GitHub GraphQL & @neo4j/graphql schemas without crashing
An enhanced Model Context Protocol (MCP) server for GraphQL that acts as a GraphQL gateway for LLMs. It is intended to handle large GitHub GraphQL and @neo4j/graphql schemas without crashing, based on real-world interoperability needs. Topics include mcp, model-context-protocol, graphql-gateway, and schema-introspection.
🛠️ Key Features
GraphQL gateway for LLMs
Enhanced MCP server for GraphQL
Designed to handle huge GitHub GraphQL schemas
Designed to handle @neo4j/graphql schemas without crashing
Emphasis on real-world interop fixes
🚀 Use Cases
Connecting MCP clients to GraphQL backends for LLM tooling
Federating or working across GraphQL schemas (as indicated by “federated-graphql” and “federation” topics)
Introspecting or supporting large GraphQL schemas (as indicated by “schema-introspection”)
⚡ Developer Benefits
More reliable schema handling for GitHub GraphQL and @neo4j/graphql
Easier MCP integration via a GraphQL gateway
Supports GraphQL and knowledge-graph oriented workflows (per topics like knowledge-graph)
⚠️ Limitations
No tool count, specific endpoints, or supported operations are provided in the available data
An enhanced MCP (Model Context Protocol) server for GraphQL that fixes real-world interoperability issues between LLMs and GraphQL APIs.
Drop-in replacement for mcp-graphql — with dynamic headers, robust variables parsing, and zero breaking changes.
🎯 What is mcp-graphql-enhanced?
mcp-graphql-enhanced is a high-performance, federated GraphQL gateway designed to act as a workhorse for LLM agents. It bridges the gap between massive, complex GraphQL ecosystems and the context-limited environment of AI assistants. Unlike standard "all-or-nothing" introspection tools that crash under the weight of large schemas (like GitHub's or enterprise-grade Neo4j graphs), this server provides surgical control over how your agent perceives and interacts with your data.
💡 Why do you need it?
If you have ever seen the<error>Tool result is too large</error>while trying to introspect your API, you are already hitting the limits of standard MCP implementations. Here is why mcp-graphql-enhanced is a practical choice for professional environments:
Avoid the 1MB Ceiling: It shifts the responsibility for scope from the server to the caller. Instead of a unilateral "everything or nothing" dump, you get granular control via typeNames and typeDepth parameters.
Surgical Precision: You can selectively introspect only the nodes you need (e.g., Repository, User, or Message), keeping your context window clean and your LLM focused.
Predictability over Immunity: It doesn't promise "unlimited" capacity—it promises predictability. In enterprise systems, you need a tool that lets you navigate the graph surgically and fail predictably if you overstep, rather than a "black box" that dies on you the moment the schema grows.
Demonstration: See a real-world diagnostic test against the GitHub API that compares scoped and monolithic introspection:
🔗 Diagnostic Case Study: Scoped vs. Monolithic Introspection (Shared Chat)
💬 Community & Support
Join the conversation! If you have questions about using this bridge with Neo4j, Discord data graphs, or GraphQL in general, come hang out with us:
This is the best place to share your feedback, report issues, or suggest new "enhanced" features for the bridge.
✨ Key Enhancements
✅ Dynamic Endpoint Switching — Hot-swap targets on the fly directly via tool arguments without restarting the server or losing session context.
✅ Built-in GraphiQL IDE — Visual playground at / (or /graphql, /graphiql) with pre-configured headers for instant testing and introspection.
✅ Dual Transport — Supports both STDIO (for local CLI/client tools) and HTTP/JSON-RPC (for external/browser clients).
✅ Dynamic headers — pass Authorization, X-API-Key, etc., via tool arguments (no config restarts)
✅ Robust variables parsing — fixes “Query variables must be a null or an object” error
✅ Smart introspection — supports filtered requests (via typeNames) and recursive depth control (via typeDepth) to minimize LLM context noise and optimize schema exploration.
✅ Full MCP compatibility — works with Claude Desktop, Groq Desktop, Google Antigravity, Glama, Gemini CLI, Hermes Agent and any standard MCP client
✅ Secure by default — mutations disabled unless explicitly enabled
✅ Dynamic Schema Evolution — Smart diagnostics and gap analysis for servers that regenerate GraphQL types on-the-fly (like Neo4j).
✅ Deep Observability — Automatic Cypher extraction and cleaning from GraphQL extensions.
🔥 Dynamic Endpoint Switching
The bridge allows LLMs or clients to dynamically target different GraphQL endpoints at runtime within a single session without requiring server restarts or configuration changes.
Simply pass the optional endpoint parameter in query-graphql or introspect-schema:
Zero Downtime: Hot-swaps the underlying schema and clears internal caches instantly.
Context Preservation: Keeps the MCP connection open while shifting queries between different environments (e.g., switching from a Discord ingest node to a Neo4j graph database).
The server operates as a Federated GraphQL Gateway capable of querying multiple independent GraphQL environments through a single endpoint.
Zero Breaking Changes: If you provide a single URL in ENDPOINT, the server behaves exactly as before.
Federated Capability Discovery: During startup, all configured endpoints are introspected to build a federated manifest used for routing, relevance analysis, and node selection.
Stable Runtime Introspection: Direct GraphQL introspection queries (__schema, __type) are always resolved against the first configured endpoint. This preserves a consistent schema model and avoids corruption caused by merging independent introspection trees.
Selective Node Routing: The federated manifest is consulted before execution to determine whether a query is relevant to a node, reducing unnecessary requests and improving performance.
Schema-Adaptive Broadcasts: Queries are automatically pruned against each participating schema before execution. Unsupported fields and empty selection branches are removed on a per-node basis, allowing a single query to operate across heterogeneous GraphQL environments.
Smart Aggregation: When multiple comma-separated URLs are provided, the server broadcasts queries, aggregates responses, and performs universal deep object-level deduplication.
Conflict Handling: Structural differences between identically named types across nodes are detected and exposed safely.
Unified Graph Experience: Independent GraphQL systems can be queried through a single endpoint, including combinations such as GitHub GraphQL, Neo4j GraphQL, and custom GraphQL APIs.
Bypass Free Tier Limits: Useful for Free Tier cloud databases (for example Neo4j Aura). Data can be distributed across multiple instances while remaining queryable through a single federated gateway.
Proof of Concept:
See a real-world demonstration of the federated query synthesis in action, where the agent aggregates live Discord data with historical Neo4j insights:
🔗 Live Federation Analysis (Shared Chat)
💡 Use Case: Bridging WSL and Windows (PowerShell)
A common challenge for Windows developers is the network isolation between the Windows Subsystem for Linux (WSL) and the host OS. This feature allows you to bridge these two worlds into a "Unified Nervous System".
Hybrid Ecosystem: Seamlessly query and aggregate data across Windows-native processes (PowerShell) and Linux-based environments (WSL).
mDNS Support: By using .local addresses, the bridge automatically resolves the host machine's IP from within the WSL environment.
Transparent Aggregation: The AI assistant interacts with a single unified schema, unaware that the data is being fetched from different operating systems simultaneously.
🔍 Advanced Observability & Cypher
The bridge provides deep insights into how the LLM interacts with your graph database.
🕸️ Automated Cypher Extraction
For GraphQL server implementations that return query execution plans (like @neo4j/graphql), the bridge automatically:
Detectsextensions.cypher in the response.
Sanitizes the output by stripping internal headers (like CYPHER 5 or empty PARAMS).
Injects a clean Cypher block directly into the tool's output for the AI to analyze.
Note: This feature requires your GraphQL server to be configured to include debug information in the response extensions.
🎨 Visual Command Center (GraphiQL)
Unlike standard MCP servers, this one provides a visual interface for humans. When running with ENABLE_HTTP=true, you can open a full-featured GraphiQL IDE in your browser.
Header Sync: Any headers set in your environment (like GitHub tokens) are automatically injected into the GraphiQL "Headers" tab for immediate testing.
💻 HTTP / Dual Transport
This server runs in dual transport mode, supporting both the standard STDIO communication (used by most MCP clients) and an HTTP JSON-RPC endpoint on port 6274.
This allows external systems, web applications, and direct curl commands to access the server's tools with live request logging in your terminal ([HTTP-RPC] logs).
Endpoint
Method
Description
/graphiql
GET
Human Interface: The visual GraphQL IDE.
/mcp
POST
The main JSON-RPC 2.0 endpoint for tool execution.
/health
GET
Simple health check, returns { status: 'ok' }.
Resolving Port Conflicts (EADDRINUSE) and Automatic Port Selection
The server defaults to port 6274. If you encounter an EADDRINUSE: address already in use :::6274 error (common in local development due to stale processes), the server will automatically find the next available port (up to 10 attempts, not spawning multiple servers).
This ensures the server starts successfully even when the default is blocked. Always check the server logs for the final bound port (e.g., [HTTP] Started server on http://localhost:6275) if your curl or client tool fails on the default 6274.
To force a specific port (e.g., for guaranteed external firewall settings), you can still explicitly set the MCP_PORT environment variable:
bash
MCP_PORT=8080 npx @letoribo/mcp-graphql-enhanced
Testing the HTTP Endpoint
You can test the endpoint using curl as long as the server is running (e.g., via npm run dev):
Test the health check (assuming the server bound to the default or found the next available port)
Example of port configuration and development mode settings.
code
# Change the HTTP port
MCP_PORT=8080 npx @letoribo/mcp-graphql-enhanced
# Test targeted introspection and explore the schema visually:
ENDPOINT=https://api.github.com/graphql \
HEADERS='{"Authorization":"Bearer YOUR_GITHUB_TOKEN"}' \
ENABLE_HTTP=true \
npx @letoribo/mcp-graphql-enhanced
# Then visit http://localhost:6274/graphiql
5. Interactive Launch with mcpgql (Recommended)
If you want to skip manual environment variable setup, use our CLI tool available on npm.
If installed globally (npm install -g @letoribo/mcpgql), simply run:
code
mcpgql
Alternatively, you can run it without global installation using:
code
npx @letoribo/mcpgql@latest
Zero-config start: Automatic discovery and template initialization.
Interactive selection: Toggle multiple endpoints and permissions on the fly.
Bridge mode: Automatically handles the Federated Bridge setup on localhost:6274.
☁️ Deploy to Cloud
This server is fully containerized and optimized for long-running processes.
Recommended Hosting
Platform
Recommended PORT
Notes
Railway
8080
Set PORT in Variables and update Networking ingress port.
Render
10000
Set PORT in Environment Variables.
Azure Container Apps
8080
Set PORT and ENABLE_HTTP=true in Environment Variables, then enable Ingress targeting the same port.
For hosted environments, you must set the environment variable ENABLE_HTTP=true in your platform's settings to ensure the HTTP transport layer is active.
Ensure your hosting provider's public networking settings are configured to route traffic to the port defined in your PORT environment variable to avoid 502 Bad Gateway errors.
⚠️ Concurrent Cloud Usage & Dynamic Endpoints
While mcp-graphql-enhanced supports switching target endpoints on the fly via the optional endpoint argument, keep state mutability in mind when running in hosted/shared environments (e.g., Render, Railway, Smithery):
State Isolation: Providing an endpoint argument dynamically updates the active graph context for the server instance.
Best Practice for Concurrent Cloud Requests: If multiple clients, agent sessions, or automated workers share the same cloud instance, always pass the endpoint explicitly in every tool call (query-graphql and introspect-schema).
Preventing Context Drift: Relying on the implicitly cached endpoint in a shared environment can lead to race conditions where a concurrent call from another session switches the active upstream URL under your feet.
🌐 Public Live Gateways (Hosted SSE/HTTP Bridges)
If you want to test the bridge instantly without running local Node.js processes, use our hosted cloud endpoints:
Note: Public gateways operate in shared environments. Remember to supply your explicit endpoint and headers in tool calls to ensure request isolation.
🖥️ Claude Desktop Configuration Examples
You can connect Claude Desktop to your GraphQL API using either the npx package (recommended for simplicity) or the Docker image (ideal for reproducibility and isolation).
Ever wished your GraphQL queries generated inside Claude Desktop or Antigravity would auto-populate directly into your GraphiQL playground in real time?
@letoribo/mcp-graphql-enhanced provides an HTTP/SSE bridge to mcp-graphiql. Whenever your AI agent calls query-graphql, the payload streams instantly into an active GraphiQL tab, while introspect-schema hot-reloads and updates the Docs explorer in real time when switching endpoints.
Resources
graphql-schema: The server exposes the GraphQL schema as a resource that clients can access. This is either the local schema file, a schema file hosted at a URL, or based on an introspection query.
Available Tools
The server provides two main tools:
introspect-schema: Retrieves the GraphQL schema or a subset. Use this first to understand the graph structure.
Arguments:
typeNames(optional, array): List of specific types to introspect (e.g., ["User", "Message"]). Reduces noise by returning only relevant parts of the graph.
typeDepth(optional, number): Controls the recursion level of nested fields (Default: 2).
Note:typeDepth is only functional when typeNames is provided to narrow the scope.
endpoint(optional, string): Target GraphQL HTTP/HTTPS URL to dynamically switch endpoint on the fly before executing introspection.
💡 Note: Dynamic Headers vs Static env
Previously, auth tokens had to be hardcoded statically at startup inside the configuration's "env" block (e.g., in claude_desktop_config.json).
With the optional headers argument added directly to query-graphql and introspect-schema, you get total dynamic control:
Keep Secrets in Your Terminal: You don't have to feed your private tokens (like GitHub PATs or Bearer keys) directly into Claude's prompt or static setup. Just set "ENABLE_HTTP": "true", run your cURL commands or local scripts directly in the terminal, and pass headers there. Any headers passed at runtime will override static tokens from your config file. Otherwise, Claude might just laugh at you for sharing your raw secret tokens in the chat!
On-the-Fly Switching: You can dynamically inject different auth headers per request or per endpoint right at execution time without restarting the MCP server.
🔒 Security & Privacy Notice
No Implicit External Traffic:
We removed the hardcoded default endpoint (previously pointing to our official demo graph endpoint, which automated security scanners flagged as an unchosen third-party host).
If no ENDPOINT environment variable is explicitly provided, the bridge no longer initiates any network requests. Instead, it securely boots into an offline local fallback mode powered by an embedded schema with built-in documentation and Tips & Tricks. Zero data leaks, zero scanner warnings.
Core Security Guidelines:
Mutations Disabled by Default: Operations that modify persistent state are blocked out-of-the-box unless ALLOW_MUTATIONS=true is explicitly set.
Input Validation: Always validate HEADERS and SCHEMA inputs in production environments.
Transport Credentials: Use HTTPS endpoints and short-lived authorization tokens where possible.
Customize for your own server
This is a very generic implementation where it allows for complete introspection and for your users to do whatever (including mutations). If you need a more specific implementation I'd suggest to just create your own MCP and lock down tool calling for clients to only input specific query fields and/or variables. You can use this as a reference.
Note on Contributors
Claude Code was not used in this project. If you're curious how Claude ended up as a co-author via custom MCP architecture, feel free to ask!
Install
Configuration
Environment variables
ENDPOINT
GraphQL endpoint(s). Supports comma-separated list for Multi-Node Architecture.
HEADERS
JSON string containing headers for requests.
ALLOW_MUTATIONS
Enable mutation operations (disabled by default).
NAME
Name of the MCP server.
SCHEMA
Path to a local GraphQL schema file or URL.
MCP_PORT
Port for the HTTP/JSON-RPC server.
ENABLE_HTTP
Enable HTTP transport: auto (default), true, or false.