io.github.aicrafted/searxng-mcp

SearXNG MCP Server

A Model Context Protocol (MCP) server that provides web search capabilities by integrating with a SearXNG instance.

Features

• Web Search: Perform powerful aggregated searches across multiple engines.
• Discovery: Programmatically retrieve available categories and engines.
• Stateless HTTP: Compatible with any standard JSON-RPC client.
• Flexible Configuration: Supports environment variables and command-line arguments.

Example of compose.yml to run SearXNG with MCP server

services:
  searxng:
    image: searxng/searxng:latest
    ports:
      - "${SEARXNG_PORT:-8080}:8080"
    volumes:
      - "${SEARXNG_VOL_CONFIG:-searxng-config}:/etc/searxng/"
      - "${SEARXNG_VOL_DATA:-searxng-data}:/var/cache/searxng/"
    restart: always

  searxng-mcp:
    image: ghcr.io/aicrafted/searxng-mcp:latest
    restart: unless-stopped
    depends_on:
      # Ensure SearXNG starts before the MCP server
      - searxng
    environment:
      SEARXNG_URL: "${SEARXNG_URL:-http://searxng:8080}"
      MCP_HOST: "${MCP_HOST:-127.0.0.1}"
      MCP_PORT: "${MCP_PORT:-32123}"
      MCP_TRANSPORT: "${MCP_TRANSPORT:-http}"
      MCP_ALLOWED_HOSTS: "${MCP_ALLOWED_HOSTS:-localhost:*,127.0.0.1:*}"
      MCP_ALLOWED_ORIGINS: "${MCP_ALLOWED_ORIGINS:-http://localhost:*,http://127.0.0.1:*}"
      MCP_DISABLE_DNS_REBINDING_PROTECTION: "${MCP_DISABLE_DNS_REBINDING_PROTECTION:-false}"
    ports:
      - "${MCP_PORT:-32123}:${MCP_PORT:-32123}"

volumes:
  searxng-config:
  searxng-data:

Important: Enable JSON responses in your SearXNG settings.yml, otherwise the MCP server cannot read search results:

search:
  formats:
    - html
    - json

Example .env

# SearXNG url should be visible by the MCP server inside docker, so use internal service port here
SEARXNG_URL=http://searxng:8080
# Public SearXNG port
SEARXNG_PORT=8080
# Searxng config and data volumes, start with "./" if You want to bind dir instead using volume
SEARXNG_VOL_CONFIG=searxng-config
SEARXNG_VOL_DATA=searxng-data

# MCP server host, port and transport ("stdio", "sse", "http")
MCP_HOST=127.0.0.1
MCP_PORT=32123
MCP_TRANSPORT=http

# MCP DNS rebinding protection (see https://github.com/modelcontextprotocol/python-sdk/issues/1798 for details)
MCP_ALLOWED_HOSTS=localhost:*,127.0.0.1:*
MCP_ALLOWED_ORIGINS=http://localhost:*,http://127.0.0.1:*
# MCP_DISABLE_DNS_REBINDING_PROTECTION=true

MCP client config

HTTP transport (recommended)

{
  "mcpServers": {
    "searxng": {
      "type": "http",
      "url": "http://localhost:32123/mcp"
    }
  }
}

SSE transport

{
  "mcpServers": {
    "searxng": {
      "type": "sse",
      "url": "http://localhost:32123/sse"
    }
  }
}

Note: SSE transport uses the /sse endpoint, not /mcp. HTTP transport uses /mcp.

Prerequisites for run from sources

• Python 3.10+
• A running SearXNG instance.

Installation

1. Clone the repository and navigate to the directory.
2. Install dependencies:

    pip install -r requirements.txt

3. Set up your .env file (optional).

Configuration

The server reads configuration from command-line arguments and environment variables. Command-line arguments override the corresponding defaults used at startup.

Usage

Run the server using uv or standard python:

python searxng_mcp.py --transport http --port 32123 --searxng http://searx.lan

Run with Docker

1. Build the image:

    docker build -t searxng-mcp .

2. Run the container:

    docker run -d \
      -p 32123:32123 \
      --env-file .env \
      --name searxng-mcp \
      searxng-mcp

Transport Options

• stdio: Standard input/output (default for some MCP clients).
• http: Stateless HTTP (streamable-http).
• sse: Server-Sent Events.

DNS Rebinding Protection

Recent versions of the MCP Python SDK validate Host and Origin headers for HTTP/SSE transports to protect local servers from DNS rebinding attacks. If you expose the server through Docker, a reverse proxy, or a custom domain and receive 421 Invalid Host Header, configure the allowlist explicitly:

MCP_ALLOWED_HOSTS=localhost:*,127.0.0.1:*,mcp.example.com:*
MCP_ALLOWED_ORIGINS=http://localhost:*,http://127.0.0.1:*,https://mcp.example.com

For trusted local development or when this validation is handled by another infrastructure layer, you can disable the SDK protection:

MCP_DISABLE_DNS_REBINDING_PROTECTION=true

Use disabling sparingly; setting MCP_ALLOWED_HOSTS and MCP_ALLOWED_ORIGINS is the recommended option.

Search Abilities Guide

SearXNG aggregates results from various sources. This guide outlines the capabilities available through the web_search tool.

Search Categories

Categories help refine your search by content type. Use these in the categories parameter (comma-separated).

Supported Engines

SearXNG can query over 130 engines. Configured engines typically include:

• Web: Google, Brave, DuckDuckGo, Qwant, Startpage
• Knowledge: Wikipedia, Wikidata
• Development: GitHub, StackOverflow, PyPI
• Social: Reddit, Twitter/X

Advanced Search Parameters

• categories: Filter by specific types (e.g., news,it).
• engines: Force specific engines (e.g., google,wikipedia).
• language: Specify search language (e.g., en, es, fr).
• pageno: Navigate through multiple pages of results.
• time_range: Filter by date (day, month, year).
• safesearch: Control content filtering (0=None, 1=Moderate, 2=Strict).

Programmatic Discovery

Use the web_search_info tool to dynamically retrieve the list of enabled categories and engines from your instance.

Windows Troubleshooting

localhost not reachable while Docker container is running

Symptom: http://localhost:<port>/ returns connection refused or hits the wrong service, but curl from inside the container works fine.

Root cause: WSL2 port relay ghost

WSL2 automatically forwards ports from the Linux VM to the Windows host using wslrelay.exe. When a process inside WSL listens on a port, WSL creates a relay bound to [::1]:<port> (IPv6 loopback) on the Windows side.

When that WSL process stops, wslrelay.exe often does not release the port. The relay entry stays alive as a zombie listener on [::1]:<port>.

Later, when Docker maps a container to the same host port, it binds correctly to 0.0.0.0:<port> — but [::1]:<port> is already taken by the stale relay.

On Windows, localhost resolves to ::1 (IPv6) first. So browser and curl requests to localhost:<port> hit the dead wslrelay.exe entry instead of the Docker container, resulting in a connection error or unexpected response.

Connecting via the explicit IPv4 address 127.0.0.1:<port> bypasses the relay and reaches Docker correctly.

How to diagnose:

# Check what is listening on the port
netstat -ano | findstr :<port>

# Identify the processes
Get-Process -Id <pid1>,<pid2> | Select-Object Id,Name

If you see two entries for the same port — one owned by com.docker.backend and another by wslrelay — this is the problem.

Workarounds:

Permanent fix:

After wsl --shutdown, restart the Docker container. The relay will no longer exist and localhost:<port> will work normally until the same port is reused inside WSL again.

Prevention:

If you regularly run services on the same port both in WSL and in Docker, prefer one of:

• Always use Docker for that service, never WSL directly
• Use different ports for WSL dev and Docker prod instances
• Add 127.0.0.1:<port>:<port> explicit binding in docker-compose.yml to force IPv4

Related

• WSL2 networking documentation
• WSL GitHub issue tracker: search wslrelay port leak