Add web search to Cursor with an MCP server

In short: Add Siftdog to Cursor’s MCP configuration for web search, page extraction, crawling and site mapping over local stdio or a running server, without a search API key.

What web tools does Cursor get?

These are read-only tools from Siftdog. You can ask Cursor Agent to use a tool by name rather than hoping it chooses one. Cursor may ask you to approve a tool call before it runs.

Run Siftdog locally in Cursor

1. Install uv so Cursor can run uvx siftdog mcp. The stdio server does not require a separately running Siftdog API server. To use search, you do need a reachable SearXNG instance with JSON output enabled; extraction, crawling and mapping work without one.

2. Create .cursor/mcp.json in your project, or use ~/.cursor/mcp.json to make the tool available across projects. Cursor supports both locations; a project entry takes priority if you use the same server name in both.

{
  "mcpServers": {
    "siftdog": {
      "command": "uvx",
      "args": ["siftdog", "mcp"],
      "env": { "SEARXNG_URL": "http://your-searxng:8080" }
    }
  }
}

Replace the example address with one Cursor’s machine can reach. If you only need extract, crawl and map, omit the env entry. Save the file and restart Cursor so it loads the configuration. Cursor documents command, args and env for local MCP servers.

Connect Cursor to a running Siftdog server

If you don’t already run SearXNG, Docker Compose is the simpler way to get search working: it starts Siftdog alongside SearXNG. See the self-hosting guide for the broader API setup.

git clone https://github.com/khsarvar/siftdog && cd siftdog
docker compose up

Replace the stdio entry in your Cursor MCP file with an HTTP entry. Use localhost for this local setup: Siftdog’s /mcp endpoint restricts which hostnames it accepts.

{
  "mcpServers": {
    "siftdog": {
      "url": "http://localhost:8000/mcp"
    }
  }
}

Cursor supports a url for remote MCP connections and a headers object when authentication is required. If you set Siftdog’s API_KEYS, add "headers": {"Authorization": "Bearer <key>"} inside the siftdog entry. Keep a file containing a real key out of a shared repository.

Check web search in Cursor

Restart Cursor, open Agent chat and ask: “Use siftdog_search to find the official Python documentation for virtual environments, then give me the URL.” You can check whether the server is enabled in Customize; Cursor lets you toggle MCP servers there.

If Cursor cannot connect, check the MCP Logs in its Output panel and confirm that the configured server is running or that uvx is available to Cursor. If extraction works but search does not, check SEARXNG_URL and whether SearXNG has JSON output enabled. If every upstream search engine fails, Siftdog reports an error rather than presenting an empty result list.

Limits

Siftdog extracts static HTML; it does not render JavaScript-heavy pages. Search also depends on public search engines through SearXNG, which may rate-limit bursts of requests. For a setup that avoids running the stack, a hosted search service may be a better fit. If you also use Claude Code, the Claude Code MCP guide covers the same Siftdog tools in that client.