generated: '2026-08-27'
method: searched
source: https://github.com/infiniflow/ragflow/blob/main/mcp/server/server.py + https://ragflow.io/docs/launch_mcp_server + https://ragflow.io/docs/mcp_tools
name: RAGFlow MCP Server
status: published
vendor: InfiniFlow Inc.
license: Apache-2.0
description: >-
  A first-party Model Context Protocol server that InfiniFlow ships inside the RAGFlow repository at
  mcp/server/server.py. It is an OPTIONAL, DISABLED-BY-DEFAULT component of a RAGFlow deployment: the
  operator uncomments services.ragflow.command in docker/docker-compose.yml (or runs
  `uv run mcp/server/server.py`) to start it alongside a working RAGFlow server. There is no
  InfiniFlow-hosted MCP endpoint — cloud.ragflow.io/mcp answers HTTP 405 from openresty, confirming no
  MCP surface is exposed on RAGFlow Cloud.
deployment:
  mode: local-stdio
  endpoint: null
  install: >-
    uv run mcp/server/server.py --host=127.0.0.1 --port=9382 --base-url=http://127.0.0.1:9380
    --api-key=<RAGFLOW_API_KEY>
  install_docker: >-
    docker compose -f docker-compose.yml up, after uncommenting services.ragflow.command with
    --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
  package: https://github.com/infiniflow/ragflow/tree/main/mcp/server
  auth: api-key
  verified: searched
  note: >-
    RECORDED AS local-stdio BECAUSE A HUMAN MUST DEPLOY IT FIRST — but the transport is not stdio. Once
    running, this server speaks MCP over HTTP: a streamable-HTTP transport at /mcp (GET/POST/DELETE,
    JSON responses on by default) and the legacy SSE transport at /sse with POSTs to /messages/. Both
    are enabled by default and each can be disabled with --no-transport-streamable-http-enabled /
    --no-transport-sse-enabled. The endpoint therefore only ever exists at an address the operator
    chooses (default 127.0.0.1:9382), never at a vendor URL, which is why `endpoint` is null.
modes:
- name: self-host
  default: true
  description: >-
    The MCP server holds one RAGFlow API key and can reach only that tenant's datasets. The key is
    supplied at launch via --api-key / --mcp-host-api-key.
- name: host
  default: false
  description: >-
    Each MCP client supplies its own RAGFlow API key on every request, so each client reaches its own
    datasets. Streamable-HTTP is not yet supported in host mode; --no-transport-streamable-http-enabled
    must be set.
requirements:
- RAGFlow server v0.18.0 or later, running and reachable at --base-url
- A RAGFlow API key (see https://ragflow.io/docs/acquire_ragflow_api_key)
tools:
- name: ragflow_retrieval
  description: >-
    Retrieve relevant chunks from the RAGFlow retrieve interface based on a question. dataset_ids may be
    omitted or empty to search across all available datasets; document_ids narrows to specific documents.
    The tool description is generated at runtime and embeds the id + description of every dataset the
    calling key can see.
  input_schema:
    type: object
    required:
    - question
    properties:
      question:
        type: string
      dataset_ids:
        type: array
        items: {type: string}
      document_ids:
        type: array
        items: {type: string}
      page:
        type: integer
        default: 1
        minimum: 1
      page_size:
        type: integer
        default: 10
        minimum: 1
        maximum: 100
      similarity_threshold:
        type: number
        default: 0.2
        minimum: 0.0
        maximum: 1.0
      vector_similarity_weight:
        type: number
        default: 0.3
        minimum: 0.0
        maximum: 1.0
      keyword:
        type: boolean
        default: false
      top_k:
        type: integer
        default: 1024
        minimum: 1
        maximum: 1024
      rerank_id:
        type: string
      force_refresh:
        type: boolean
        default: false
- name: ragflow_list_datasets
  description: >-
    List all accessible datasets (knowledge bases), returning dataset ids, names and descriptions, so a
    client can discover what to retrieve from.
  input_schema:
    type: object
    properties:
      page:
        type: integer
        default: 1
        minimum: 1
      page_size:
        type: integer
        default: 100
        minimum: 1
        maximum: 1000
- name: ragflow_list_chats
  description: >-
    List all accessible chat assistants, returning ids, names and descriptions, so a client can discover
    which assistants are available for conversation.
  input_schema:
    type: object
    properties:
      page:
        type: integer
        default: 1
        minimum: 1
      page_size:
        type: integer
        default: 30
        minimum: 1
        maximum: 100
tool_count: 3
docs_note: >-
  https://ragflow.io/docs/mcp_tools still documents only a single `retrieve` tool. The shipped server
  source exposes three tools named ragflow_retrieval, ragflow_list_datasets and ragflow_list_chats — the
  published docs page is behind the code. Schemas above are read verbatim from server.py, not from the
  docs.
x-evidence:
- url: https://cloud.ragflow.io/mcp
  method: POST tools/list
  http_status: 405
  note: openresty "405 Not Allowed" — RAGFlow Cloud exposes no MCP endpoint.
- url: https://raw.githubusercontent.com/infiniflow/ragflow/main/mcp/server/server.py
  http_status: 200
  note: 38,665 bytes; tool definitions read at lines 542-649.