SearXNG

SearXNG MCP Server

A Model Context Protocol (MCP) server that equips AI agents with privacy-friendly web search capabilities using the SearXNG metasearch engine.

Features

This server provides the following main features:

• Web search via multiple search engines
• Supports various search categories (general, images, news, etc.)
• Customizable search engine selection
• Language filtering
• Time range filtering
• Control over the number of search results

Available Tools

• web_search - Perform web search using SearXNG
  • Required parameters:
    • query (string): The search query
  • Optional parameters:
    • categories (array): Search categories, e.g. ['general', 'images', 'news']
    • engines (array): Search engines, e.g. ['google', 'bing', 'duckduckgo']
    • language (string): Language code for search, default is "en"
    • max_results (integer): Maximum number of results, default is 10 (1-100)
    • time_range (string): Time range filter ('day', 'week', 'month', 'year')

If a search cannot be completed (the instance is unreachable, rate-limits the request, or returns a malformed response), the tool returns an error result describing the failure rather than an empty result list.

Search Limits

Each server process admits up to eight concurrent searches. Further calls return an error immediately and can be retried later. --timeout bounds the complete network operation, including streaming the response; cancellation closes the active request. Responses are limited to 2 MiB before JSON parsing.

The configured instance must serve /search directly: redirects are rejected. The client requests Accept-Encoding: identity and rejects compressed responses to prevent unbounded decompression. Result fields are also truncated to their existing limits. Missing or non-list results and upstream error objects are reported as failures, while a valid empty list remains a successful search.

Application logs omit search queries and upstream exception text. The CLI keeps HTTP dependency logging at WARNING even when --log-level=DEBUG is selected.

Privacy Policy

This software does not operate a hosted search service and does not send search queries to its maintainer. It runs on the user's device or infrastructure and sends each query, selected engines, categories, language, time range, and safe search setting directly to the configured SearXNG instance. Search results are returned to the connected MCP client.

The software does not persist queries or results, use analytics or advertising trackers, create user profiles, or share data with the maintainer. Operational logs record result counts and errors but omit query text and upstream exception details. In-memory request and result data is discarded after the request.

The selected SearXNG instance and the search engines it contacts are independent third parties. Their collection, use, sharing, and retention practices are governed by their own privacy policies. Before sending sensitive queries, users should review the selected instance's privacy notice or operate an instance they trust. The default CLI instance is https://searx.party; its policy is at https://searx.party/info/en/privacy.

Privacy questions and requests can be filed at https://github.com/danchev/searXNG/issues.

Command Line Options

Option	Default	Description
--instance-url	https://searx.party	SearXNG instance base URL. Must be absolute and cannot contain credentials, a query string, or a fragment.
--timeout	30	Total search network timeout, in seconds.
--log-level	WARNING	Logging verbosity: DEBUG, INFO, WARNING, ERROR, CRITICAL. Logs are written to stderr.
--transport	stdio	Transport to serve on: stdio or http.
--host	127.0.0.1	Host to bind when --transport=http.
--port	8000	Port to bind when --transport=http.
--header	None	Custom HTTP headers to send with search requests (e.g., X-Forwarded-For: 1.2.3.4). Can be specified multiple times.

Docker / Kubernetes

You can run the server easily using the official container image published to the GitHub Container Registry:

docker run -p 8000:8000 ghcr.io/danchev/searxng:latest \
    --transport http \
    --host 0.0.0.0 \
    --instance-url=https://searx.party \
    --header "Authorization: Bearer my-secret-token"

Transports

By default the server speaks stdio, which is what local MCP clients (Claude Desktop, IDE integrations, uvx) launch it with.

For remote access, --transport http serves the Streamable HTTP transport at /mcp:

searxng --transport http --host 127.0.0.1 --port 8000
# endpoint: http://127.0.0.1:8000/mcp

The legacy SSE transport is intentionally not implemented. It was superseded by Streamable HTTP in the 2025-03-26 MCP protocol revision and should not be used for new deployments.

Security: the server performs no authentication, so it binds to 127.0.0.1 by default. Only pass --host 0.0.0.0 on a trusted network, or put an authenticating reverse proxy in front of it.

Binding to a non-loopback host also disables the MCP SDK's DNS-rebinding protection, which it can only enable automatically for 127.0.0.1, localhost, and ::1. On a loopback bind a forged Host header is rejected with 421 Misdirected Request; on a public bind any Host is accepted. The server logs a warning at startup when this applies.

Usage Example

Configure as an MCP Service

To set up SearXNG as an MCP server, add one of the following to your MCP configuration file:

UVX setup:

"mcpServers": {
  "searxng": {
    "command": "uvx",
    "args": ["searxng", "--instance-url=https://searx.party"]
  }
}

Docker setup (Local stdio):

"mcpServers": {
  "searxng": {
    "command": "docker",
    "args": [
      "run",
      "-i",
      "--rm",
      "ghcr.io/danchev/searxng:latest",
      "--instance-url=https://searx.party"
    ]
  }
}

This launches the server over stdio, which is the right choice for a local client.

Remote setup (Streamable HTTP):

Start the server as a long-running process:

searxng --transport http --host 0.0.0.0 --port 8000 \
        --instance-url=https://searx.party

Then point the client at its /mcp endpoint:

"mcpServers": {
  "searxng": {
    "url": "http://your-host:8000/mcp"
  }
}

Note the --host 0.0.0.0 needed to accept connections from other machines, and the security caveat above: the server is unauthenticated, so restrict it to a trusted network or front it with an authenticating reverse proxy.

Example Invocation

{
  "name": "web_search",
  "arguments": {
    "query": "climate change research",
    "categories": ["general"],
    "engines": ["google"],
    "language": "en",
    "max_results": 15,
    "time_range": "month"
  }
}

Debugging

You can use the MCP inspector to debug the server:

npx @modelcontextprotocol/inspector uvx searxng

License

AGPLv3+ License - see LICENSE for details.