Skip to main content
Glama
ragifix-mcp

ragifix-mcp

Official
by hephaistools

ragifix-mcp

MCP server (Model Context Protocol, Streamable HTTP transport) that exposes a ragifix database to a compatible chat client (search, addition, deletion, and listing of documents).

Table of contents


Related MCP server: mcp-rag-bridge

Installation via Docker

git clone <url-du-dépôt> ragifix-mcp && cd ragifix-mcp

cp config.example.yaml config.yaml              # puis l'adapter
cp deploy/ragifix-mcp.env.example ragifix-mcp.env # puis renseigner les secrets

docker build -f deploy/docker/Dockerfile -t ragifix-mcp .

docker run -d \
  --name ragifix-mcp \
  --network host \
  --env-file ragifix-mcp.env \
  -v "$(pwd)/config.yaml:/etc/ragifix/config.yaml:ro" \
  ragifix-mcp

--network host: by default (server.host: 127.0.0.1), the server must remain reachable locally on the machine that also hosts ragifix and the chat client.

Verify that the service is running (401 expected without a token — this is the sign that the server is responding):

curl -s -o /dev/null -w "%{http_code}\n" http://127.0.0.1:8422/mcp \
  -X POST -H "Content-Type: application/json" -d '{}'

Installation via .deb package

git clone <url-du-dépôt> ragifix-mcp && cd ragifix-mcp

sudo apt-get install -y devscripts debhelper python3-venv python3-pip
dpkg-buildpackage -us -uc -b

sudo apt install -y ../ragifix-mcp_0.1.0-1_all.deb

The installation (postinst) builds a Python virtual environment in /opt/ragifix-mcp/venv and installs the dependencies from PyPI (lightweight: mcp, httpx, starlette, uvicorn, pydantic — network access required at apt install time, the service then works offline).

The package creates a dedicated system user, /etc/ragifix-mcp/config.yaml and ragifix-mcp.env (from the examples), and a systemd unit ragifix-mcp.service.

sudo nano /etc/ragifix-mcp/config.yaml
sudo nano /etc/ragifix-mcp/ragifix-mcp.env
sudo systemctl enable --now ragifix-mcp
journalctl -u ragifix-mcp -f
sudo apt remove ragifix-mcp    # conserve la config
sudo apt purge ragifix-mcp     # supprime tout

Installation in a development environment

git clone <url-du-dépôt> ragifix-mcp && cd ragifix-mcp

python3 -m venv venv
source venv/bin/activate
pip install -e .

cp config.example.yaml config.yaml   # reste à côté du code, hors /etc

export RAGIFIX_MCP_TOKEN=dev-mcp-token
export RAGIFIX_API_TOKEN=dev-token    # le token de l'API ragifix visée

ragifix-mcp --config ./config.yaml

Configuration to paste into the chat client

Once the server is started, declare a remote MCP server (HTTP transport) in the client configuration, in the current standard format:

{
  "mcpServers": {
    "ragifix": {
      "url": "http://127.0.0.1:8422/mcp",
      "headers": {
        "Authorization": "Bearer <RAGIFIX_MCP_TOKEN>"
      }
    }
  }
}

Replace <RAGIFIX_MCP_TOKEN> with the value configured in ragifix-mcp.env (or RAGIFIX_MCP_TOKEN in development), and the URL with the actual server address if different from 127.0.0.1:8422.

Some older clients do not yet know how to connect directly to a remote HTTP server and only accept a local command (command/args); in that case, use a bridge such as mcp-remote:

{
  "mcpServers": {
    "ragifix": {
      "command": "npx",
      "args": [
        "-y", "mcp-remote",
        "http://127.0.0.1:8422/mcp",
        "--header", "Authorization: Bearer <RAGIFIX_MCP_TOKEN>"
      ]
    }
  }
}

Refer to the documentation of the chat client used to know where to paste this configuration (MCP config file, or management interface for MCP connectors/servers).

Exposed tools

Tool

Description

rag_query(query, top_k=5)

Searches for the most relevant passages.

rag_list_documents(prefix=None)

Lists the documents indexed in ragifix.

rag_get_document(doc_id)

Details of a document.

rag_add_document(doc_id, content, extension, content_encoding="utf8", metadata=None)

Adds/updates a document (content_encoding: "utf8" for plain text, "base64" for binary content — pdf/docx/pptx/xlsx/html).

rag_delete_document(doc_id)

Deletes a document (idempotent).

rag_health()

Checks that ragifix is reachable.

Remote access

By default (server.host: 127.0.0.1), the server is only reachable from the local machine — suitable for a chat client running on the same machine (e.g., Claude Desktop locally).

For a remote chat client, two changes are required:

  1. server.host in config.yaml: set it to 0.0.0.0 (all interfaces) or to the IP of the relevant network interface. A warning is logged at startup as long as host is not local, to remind you of this point.

  2. TLS: this server does not terminate TLS itself. Place a reverse proxy (nginx, Caddy, Traefik...) in front of it, which handles the certificate and relays to http://127.0.0.1:<port>/mcp internally. Minimal example with Caddy:

    mcp.exemple.fr {
        reverse_proxy 127.0.0.1:8422
    }

The url declared on the chat client side then becomes https://mcp.exemple.fr/mcp.

The dedicated token (RAGIFIX_MCP_TOKEN) remains the application-level protection in both cases — TLS protects the transport, the token protects access.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      D
      maintenance
      Exposes an existing RAG API as MCP tools, enabling health checks and document-based question answering with source evidence.
      -
    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables AI agents to query and manage a document knowledge base via MCP, with RAG-powered search and grounded answers with citations.
      MIT
    • F
      license
      A
      quality
      D
      maintenance
      MCP bridge to a multimodal RAG service, enabling hybrid search and Q&A over documents with tools for knowledge base queries and health checks.
      4
      -
    • A
      license
      Not graded
      quality
      C
      maintenance
      Exposes a markdown knowledge base to MCP clients so they can search pages, fetch individual documents, and ask grounded questions that come back with numbered citations or a refusal. Retrieval is graded before answering, with a single query rewrite before giving up when nothing relevant is found.
      1
      MIT