RAGFlow MCP Server for Claude

Overview

RAGFlow includes an MCP server that runs as an optional component next to a working RAGFlow deployment. It gives MCP clients such as Claude access to RAGFlow retrieval over selected datasets, using RAGFlow's DeepDoc-powered document parsing and retrieval pipeline.

The current MCP surface is intentionally focused: it exposes retrieval over dataset_ids and optional document_ids for a user-provided question. That makes it a good fit for agents that need grounded chunks from an existing RAGFlow knowledge base without giving the agent broad control over the rest of the RAGFlow application.

How It Works

The RAGFlow MCP server is designed as an independent component that complements the main RAGFlow server. It must run alongside a functioning RAGFlow deployment.

RAGFlow documents two launch modes:

The server supports the legacy SSE endpoint at /sse and streamable HTTP at /mcp. RAGFlow notes that host mode currently requires disabling streamable HTTP, so prefer self-host mode for streamable HTTP client setups unless your deployment has tested a different configuration.

Tool

RAGFlow currently documents one specialized MCP retrieval tool:

The tool description includes available dataset IDs and descriptions so the client can decide which datasets are relevant to a question.

Launch From Source

Run from a RAGFlow checkout with a working RAGFlow server:

uv run mcp/server/server.py \
  --host=127.0.0.1 \
  --port=9382 \
  --base-url=http://127.0.0.1:9380 \
  --api-key=ragflow-xxxxx

Host mode is available when each MCP client should provide its own key:

uv run mcp/server/server.py \
  --host=127.0.0.1 \
  --port=9382 \
  --base-url=http://127.0.0.1:9380 \
  --mode=host

Launch With Docker

RAGFlow's Docker workflow requires enabling the MCP server command in docker/docker-compose.yml. The documented configuration includes:

command:
  - --enable-mcpserver
  - --mcp-host=0.0.0.0
  - --mcp-port=9382
  - --mcp-base-url=http://127.0.0.1:9380
  - --mcp-script-path=/ragflow/mcp/server/server.py
  - --mcp-mode=self-host
  - --mcp-host-api-key=ragflow-xxxxxxx

For safer local development, bind the MCP host to 127.0.0.1 unless another host is explicitly required and protected.

Client Configuration

For streamable HTTP clients, point the MCP client at the local /mcp endpoint:

{
  "mcpServers": {
    "ragflow": {
      "type": "http",
      "url": "http://127.0.0.1:9382/mcp",
      "headers": {
        "Authorization": "YOUR_RAGFLOW_API_KEY"
      }
    }
  }
}

For legacy SSE clients, RAGFlow's client examples use http://127.0.0.1:9382/sse and pass the API key in either an api_key header or an Authorization header.

Use Cases

• Let Claude answer implementation questions from a private RAGFlow dataset.
• Retrieve chunks from a curated support, legal, or operations knowledge base.
• Query only approved dataset IDs instead of exposing every document in a RAGFlow deployment.
• Test dataset chunking and retrieval quality before wiring RAGFlow into broader agent workflows.
• Keep RAG search local to an existing self-hosted RAGFlow stack.

Safety and Privacy

RAGFlow MCP is retrieval-focused, but the retrieved content can still be highly sensitive. Scope dataset IDs deliberately, keep API keys outside prompts, and avoid exposing the MCP server on public interfaces without additional network controls. When using retrieved chunks in generated answers, preserve citations and verify that the dataset actually supports the claim.

Source Verification Notes

Verified on 2026-06-18:

• RAGFlow's MCP launch documentation confirms source and Docker launch paths, self-host and host modes, v0.18.0+ prerequisite, API-key behavior, /sse and /mcp transports, and network-exposure guidance.
• RAGFlow's MCP tools documentation confirms the retrieval-focused MCP tool for fetching relevant chunks from specified dataset IDs and optional document IDs.
• RAGFlow's MCP client examples document API-key headers, SSE initialization, tool listing, and a ragflow_retrieval tool call.
• The upstream infiniflow/ragflow repository is Apache-2.0 licensed and contains the mcp implementation directory.