io.github.abahocodes/llmgraph

@llmgraph/mcp-server

A Model Context Protocol (MCP) server that exposes your LLMGraph workflow deployments as MCP tools. Connect it to Claude Desktop, Claude Code, Cursor, or any other MCP client, and your assistant can invoke the workflows you built and deployed on LLMGraph.

Each configured deployment becomes one MCP tool. The server runs over stdio and is designed to be launched with npx, so there is nothing to install permanently.

Prerequisites

• Node.js 18 or newer
• A deployed LLMGraph workflow: copy the deployment endpoint URL (shaped like https://llmgraph.ai/api/<graph_id>/<environment>) and an API key from the LLMGraph dashboard

Configuration

All configuration is via environment variables.

Single deployment (simple path)

Variable	Required	Description
LLMGRAPH_ENDPOINT	yes	Full deployment endpoint URL copied from the dashboard
LLMGRAPH_API_KEY	yes	Secret API key for the deployment
LLMGRAPH_TOOL_NAME	no	Tool name shown to the client (default: invoke_workflow)
LLMGRAPH_TOOL_DESCRIPTION	no	Tool description shown to the model
LLMGRAPH_SCHEMA_MODE	no	input (default) or chat, see below
LLMGRAPH_TIMEOUT_MS	no	Request timeout in milliseconds, positive integer (default: 180000). Applies in both single and multiple deployment modes.

Multiple deployments (advanced path)

Set LLMGRAPH_DEPLOYMENTS to a JSON array; each entry becomes one tool. When set, it takes precedence over the single-deployment variables.

[
  {
    "name": "summarize_document",
    "description": "Summarizes a document with the LLMGraph summarizer workflow",
    "endpoint": "https://llmgraph.ai/api/abc123/production",
    "apiKey": "your-api-key"
  },
  {
    "name": "support_bot",
    "description": "Asks the support assistant workflow a question",
    "endpoint": "https://llmgraph.ai/api/def456/production",
    "apiKey": "your-other-api-key",
    "inputSchema": "chat"
  }
]

Schema modes

• input (default): the tool takes { "input": <object> } and the object is passed through unchanged as the POST body, so it works with any workflow input shape.
• chat: for chat-style workflows. The tool takes { "user_input": <string>, "history": [{"role": "user"|"assistant", "content": <string>}] } (history optional) and sends it in the shape chat workflows expect.

Client setup

Claude Desktop

Add to claude_desktop_config.json (Settings, Developer, Edit Config):

{
  "mcpServers": {
    "llmgraph": {
      "command": "npx",
      "args": ["-y", "@llmgraph/mcp-server"],
      "env": {
        "LLMGRAPH_ENDPOINT": "https://llmgraph.ai/api/abc123/production",
        "LLMGRAPH_API_KEY": "your-api-key",
        "LLMGRAPH_TOOL_NAME": "summarize_document",
        "LLMGRAPH_TOOL_DESCRIPTION": "Summarizes a document with my LLMGraph workflow"
      }
    }
  }
}

Restart Claude Desktop and the tool appears in the tools menu.

Claude Code

claude mcp add llmgraph \
  --env LLMGRAPH_ENDPOINT=https://llmgraph.ai/api/abc123/production \
  --env LLMGRAPH_API_KEY=your-api-key \
  -- npx -y @llmgraph/mcp-server

Cursor

Add to ~/.cursor/mcp.json (or .cursor/mcp.json in your project):

{
  "mcpServers": {
    "llmgraph": {
      "command": "npx",
      "args": ["-y", "@llmgraph/mcp-server"],
      "env": {
        "LLMGRAPH_ENDPOINT": "https://llmgraph.ai/api/abc123/production",
        "LLMGRAPH_API_KEY": "your-api-key"
      }
    }
  }
}

Error handling

Non-200 responses from the LLMGraph API are returned to the client as MCP tool errors carrying the API's error message:

Status	Meaning
400	invalid request body
401	missing or invalid API key
402	subscription blocked
403	API disabled or origin not allowed
404	unknown deployment or wrong API key
422	workflow run failed
429	rate or budget limited
504	workflow timed out

Security notes

• LLMGraph API keys are secrets for server-side use. This server sends the key only as the x-api-key header of requests to your configured endpoint, and never writes it to stdout, stderr, or error messages.
• Client config files like claude_desktop_config.json store the key in plain text on your machine; treat them accordingly.

Development

npm install
npm run build   # compiles TypeScript to dist/
npm test        # builds, then runs unit tests (node --test), no network calls

License

MIT, see LICENSE.