Search tools are easy to find; ones that hand back grounding metadata are the ones whose output you can check. Because it bills against your own Gemini API key, search volume lands on that account rather than a separate search vendor. Take the API key path — the OAuth alternative is labelled experimental and drags in a billing-enabled Cloud project.
A single-tool MCP server lifted from Google's official Gemini CLI. It runs a Google web search through the Gemini API and returns results with grounding metadata, so answers carry source links rather than arriving unattributed.
- `google_web_search` — takes a `query` string and returns search results with citations and source links
- Grounding metadata alongside the results, which is what makes the sources checkable
- A choice of model through `GEMINI_MODEL`, defaulting to `gemini-2.0-flash-exp`
Node.js 18+ and a Google API key with Gemini API access, minted in Google AI Studio and passed as `GOOGLE_API_KEY` or `GEMINI_API_KEY`. The package is `gemini-google-web-search-mcp` (1.0.0 in package.json); the documented path is a clone plus `npm install` and `npm run build`, then pointing the client at `dist/index.js`. An OAuth route via Google's Code Assist API exists behind `USE_OAUTH`, but the README marks it experimental and not fully supported, and it additionally wants a Google Cloud project with billing, the Cloud AI Companion API enabled and a Gemini Code Assist subscription. Apache 2.0.