sidebar_position: 4

slug: /mcp_server

RAGFlow MCP server overview

The RAGFlow Model Context Protocol (MCP) server operates as an independent component that complements the RAGFlow server. However, it requires a RAGFlow server to work functionally well, meaning, the MCP client and server communicate with each other in MCP HTTP+SSE mode (once the connection is established, server pushes messages to client only), and responses are expected from RAGFlow server.

The MCP server currently offers a specific tool to assist users in searching for relevant information powered by RAGFlow DeepDoc technology:

retrieve: Fetches relevant chunks from specified dataset_ids and optional document_ids using the RAGFlow retrieve interface, based on a given question. Details of all available datasets, namely, id and description, are provided within the tool description for each individual dataset.

Launching the MCP Server

Similar to launching the RAGFlow server, the MCP server can be started either from source code or via Docker.

Launch Modes

The MCP server supports two launch modes:

Self-Host Mode:
- In this mode, the MCP server is launched to access a specific tenant’s datasets.
- This is the default mode.
- The --api_key argument is required to authenticate the server with the RAGFlow server.
- Example: bash uv run mcp/server/server.py --host=127.0.0.1 --port=9382 --base_url=http://127.0.0.1:9380 --mode=self-host --api_key=ragflow-xxxxx
Host Mode:
- In this mode, the MCP server allows each user to access their own datasets.
- To ensure secure access, a valid API key must be included in the request headers to identify the user.
- The --api_key argument is not required during server launch but must be provided in the headers on each client request for user authentication.
- Example: bash uv run mcp/server/server.py --host=127.0.0.1 --port=9382 --base_url=http://127.0.0.1:9380 --mode=host

Launching from Source Code

All you need to do is stand on the right place and strike out command, assuming you are on the project working directory.

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

For testing purposes, there is an MCP client example provided, free to take!

Required Arguments

host: Specifies the server’s host address.
port: Defines the server’s listening port.
base_url: The address of the RAGFlow server that is already running and ready to handle tasks.
mode: Launch mode, only accept self-host or host.
api_key: Required when mode is self-host to authenticate the MCP server with the RAGFlow server.

Here are three augments required, the first two,host and port, are self-explained. Thebase_url is the address of the ready-to-serve RAGFlow server to actually perform the task.

Launching from Docker

Building a standalone MCP server image is straightforward and easy, so we just proposed a way to launch it with RAGFlow server here.

Alongside RAGFlow

As MCP server is an extra and optional component of RAGFlow server, we consume that not everybody going to use it. Thus, it is disable by default. To enable it, simply find docker/docker-compose.yml to uncomment services.ragflow.command section.

services:
  ragflow:
    ...
    image: ${RAGFLOW_IMAGE}
    # example to setup MCP server
    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 # `self-host` or `host`
      - --mcp-host-api-key=ragflow-xxxxxxx # only need to privide when mode is `self-host` and use bare string without quotation marks here.

To troubleshoot, launch the service in the foreground using docker compose -f docker-compose.yml.

For those upgrading from versions before v0.18.0

Get all MCP related files ready.
1. copy mcp/ directory to local.
2. copy docker/docker-compose.yml to local.
3. copy docker/entrypoint.sh to local.
4. resolve necessary dependencies via uv.
  - simply run uv add mcp if it works for you. Or:
  - copy pyproject.toml and run uv sync --python 3.10 --all-extras.
Change docker-compose.yml to enable MCP as it is disable by default, see last section.
Launch the service with docker compose -f docker-compose.yml up -d

Check the MCP server status

Checking logs of RAGFlow server with docker logs ragflow-server. If you see the MCP server ASCII art there, it means all is OK!

ragflow-server  | Starting MCP Server on 0.0.0.0:9382 with base URL http://127.0.0.1:9380...
ragflow-server  | Starting 1 task executor(s) on host 'dd0b5e07e76f'...
ragflow-server  | 2025-04-18 15:41:18,816 INFO     27 ragflow_server log path: /ragflow/logs/ragflow_server.log, log levels: {'peewee': 'WARNING', 'pdfminer': 'WARNING', 'root': 'INFO'}
ragflow-server  | 
ragflow-server  | __  __  ____ ____       ____  _____ ______     _______ ____
ragflow-server  | |  \/  |/ ___|  _ \     / ___|| ____|  _ \ \   / / ____|  _ \
ragflow-server  | | |\/| | |   | |_) |    \___ \|  _| | |_) \ \ / /|  _| | |_) |
ragflow-server  | | |  | | |___|  __/      ___) | |___|  _ < \ V / | |___|  _ <
ragflow-server  | |_|  |_|\____|_|        |____/|_____|_| \_\ \_/  |_____|_| \_\
ragflow-server  |     
ragflow-server  | MCP launch mode: self-host
ragflow-server  | MCP host: 0.0.0.0
ragflow-server  | MCP port: 9382
ragflow-server  | MCP base_url: http://127.0.0.1:9380
ragflow-server  | INFO:     Started server process [26]
ragflow-server  | INFO:     Waiting for application startup.
ragflow-server  | INFO:     Application startup complete.
ragflow-server  | INFO:     Uvicorn running on http://0.0.0.0:9382 (Press CTRL+C to quit)
ragflow-server  | 2025-04-18 15:41:20,469 INFO     27 found 0 gpus
ragflow-server  | 2025-04-18 15:41:23,263 INFO     27 init database on cluster mode successfully
ragflow-server  | 2025-04-18 15:41:25,318 INFO     27 load_model /ragflow/rag/res/deepdoc/det.onnx uses CPU
ragflow-server  | 2025-04-18 15:41:25,367 INFO     27 load_model /ragflow/rag/res/deepdoc/rec.onnx uses CPU
ragflow-server  |         ____   ___    ______ ______ __               
ragflow-server  |        / __ \ /   |  / ____// ____// /____  _      __
ragflow-server  |       / /_/ // /| | / / __ / /_   / // __ \| | /| / /
ragflow-server  |      / _, _// ___ |/ /_/ // __/  / // /_/ /| |/ |/ / 
ragflow-server  |     /_/ |_|/_/  |_|\____//_/    /_/ \____/ |__/|__/                             
ragflow-server  | 
ragflow-server  |     
ragflow-server  | 2025-04-18 15:41:29,088 INFO     27 RAGFlow version: v0.18.0-285-gb2c299fa full
ragflow-server  | 2025-04-18 15:41:29,088 INFO     27 project base: /ragflow
ragflow-server  | 2025-04-18 15:41:29,088 INFO     27 Current configs, from /ragflow/conf/service_conf.yaml:
ragflow-server  |  ragflow: {'host': '0.0.0.0', 'http_port': 9380}
...
ragflow-server  |  * Running on all addresses (0.0.0.0)
ragflow-server  |  * Running on http://127.0.0.1:9380
ragflow-server  |  * Running on http://172.19.0.6:9380
ragflow-server  |   ______           __      ______                     __            
ragflow-server  |  /_  __/___ ______/ /__   / ____/  _____  _______  __/ /_____  _____
ragflow-server  |   / / / __ `/ ___/ //_/  / __/ | |/_/ _ \/ ___/ / / / __/ __ \/ ___/
ragflow-server  |  / / / /_/ (__  ) ,<    / /____>  </  __/ /__/ /_/ / /_/ /_/ / /    
ragflow-server  | /_/  \__,_/____/_/|_|  /_____/_/|_|\___/\___/\__,_/\__/\____/_/                               
ragflow-server  |     
ragflow-server  | 2025-04-18 15:41:34,501 INFO     32 TaskExecutor: RAGFlow version: v0.18.0-285-gb2c299fa full
ragflow-server  | 2025-04-18 15:41:34,501 INFO     32 Use Elasticsearch http://es01:9200 as the doc engine.
...

You are ready to brew🍺!

Testing and Usage

Typically, there are various ways to utilize an MCP server. You can integrate it with LLMs or use it as a standalone tool. You find the way.

Example MCP Client

#
#  Copyright 2025 The InfiniFlow Authors. All Rights Reserved.
#
#  Licensed under the Apache License, Version 2.0 (the "License");
#  you may not use this file except in compliance with the License.
#  You may obtain a copy of the License at
#
#      http://www.apache.org/licenses/LICENSE-2.0
#
#  Unless required by applicable law or agreed to in writing, software
#  distributed under the License is distributed on an "AS IS" BASIS,
#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#  See the License for the specific language governing permissions and
#  limitations under the License.
#


from mcp.client.session import ClientSession
from mcp.client.sse import sse_client


async def main():
    try:
        # To access RAGFlow server in `host` mode, you need to attach `api_key` for each request to indicate identification.
        # async with sse_client("http://localhost:9382/sse", headers={"api_key": "ragflow-IyMGI1ZDhjMTA2ZTExZjBiYTMyMGQ4Zm"}) as streams:
        async with sse_client("http://localhost:9382/sse") as streams:
            async with ClientSession(
                streams[0],
                streams[1],
            ) as session:
                await session.initialize()
                tools = await session.list_tools()
                print(f"{tools.tools=}")
                response = await session.call_tool(name="ragflow_retrieval", arguments={"dataset_ids": ["ce3bb17cf27a11efa69751e139332ced"], "document_ids": [], "question": "How to install neovim?"})
                print(f"Tool response: {response.model_dump()}")

    except Exception as e:
        print(e)


if __name__ == "__main__":
    from anyio import run

    run(main)

Security and Concerns

Since MCP technology is still in booming age and there are still no official Authentication and Authorization best practices to follow, RAGFlow uses api_key to validate the identification, and it is required to perform any operations mentioned in the preview section. Obviously, this is not a premium solution to do so, thus this RAGFlow MCP server is not expected to exposed to public use as it could be highly venerable to be attacked. For local SSE server, bind only to localhost (127.0.0.1) instead of all interfaces (0.0.0.0). For additional guidance, you can refer to MCP official website.

mcp.md 10KB Originale Blame Cronologia