flexible-graphrag

Flexible GraphRAG includes an MCP server that exposes a set of tools for working with documents, data sources, and AI-powered queries via Claude Desktop MCP clients and other MCP-enabled tools. The MCP server provides full API parity for its available actions, including ingesting documents from all 13 data sources, performing hybrid searches across vector, full-text, and graph databases, and conducting AI queries and chats. It supports both HTTP mode for debugging and stdio mode for production, enabling flexible integration into your existing pipelines. The MCP client tools cover ingest_documents(), ingest_text(), search_documents(), query_documents(), system diagnostics, and health checks, along with specialized tools for processing and management. You can choose between running the MCP server directly via uvx (no-install) or installing via pipx for a persistent environment. The server works in tandem with the FastAPI backend and the UI clients (Angular, React, Vue) that consume the REST APIs provided by the FastAPI layer.

Prerequisites:

• Python 3.8+ (for MCP server components if using Python-based execution)
• uvx (no-install execution) or pipx (installation-based)
• Optional: Docker for modular deployment (docker-compose includes vector, graph, search databases, and Alfresco)

Installation steps (choose one execution path):

Option A: No-install (uvx)

1. Ensure uvx is installed on your system.
2. Run the MCP server in no-install mode: uvx flexible-graphrag-mcp
3. Verify the server starts and exposes the MCP endpoints at the configured HTTP transport or stdio channel.

Option B: Persistent environment (pipx)

1. Install the MCP package with pipx: pipx install flexible-graphrag-mcp
2. Run the MCP server via pipx: pipx run flexible-graphrag-mcp
3. Confirm the server is up and listening on the expected ports.

Option C: Docker deployment (recommended for modular databases and quick start)

1. Navigate to the docker/ directory and start the stack with docker-compose: cd docker docker-compose up -d
2. The MCP server will be available through the configured API endpoints while databases (vector, graph, search) are brought up via the compose file.

Notes:

• The MCP server supports both HTTP mode for debugging and stdio mode for production; choose the invocation mode that fits your workflow.
• If you are using Docker, you can enable or disable components (vector, graph, search engines, Alfresco) with single-line comments in the docker-compose file.

Tips and common issues:

• Environment variables: Set FGR_MCP_API_BASE_URL to your MCP REST API base URL if using HTTP transport; adjust LOG_LEVEL for troubleshooting; provide DATABASE_URL if your backend requires a separate connection string.
• Data sources: The MCP client tools support all 13 data sources. Use ingest_documents() with source-specific configs (filesystem, Alfresco, SharePoint, Box, CMIS, cloud storage, web, etc.). The skip_graph flag can disable GraphRAG processing for a given ingestion.
• Ingestion parameters: You can pass paths for filesystem/Alfresco/CMIS and, for Alfresco, nodeDetails for multi-select KG Spaces.
• Diagnostics: Use system diagnostics and health checks frequently when deploying to new environments or when you add new data sources or databases.
• Deployment mode: If you’re aiming for rapid experimentation, start with the HTTP/stdio dual transport in development, then move to a full Docker deployment for production stability.