pi-gemini-search

Web search & research tools for the Pi coding agent, backed by the Gemini API with Google Search grounding. The agent can call them autonomously to pull fresh information from the web — without dumping raw page contents into your conversation.

The key idea: search context stays inside the tool result. Gemini does the searching and reading, then returns only a synthesized answer plus source URLs, so your main context stays lean. No headless browser, no page-scraping noise.

Tools

Registers two tools the model can call:

• web_search — quick fact-check / verification. Returns a concise synthesized answer (a few sentences) + source URLs. Use for version numbers, dates, single facts, a quick second opinion. Uses url_context too, so you can hand it URLs to cross-check.
• web_research — in-depth research on a complex topic. Returns a thorough, structured answer + source URLs. Use for multi-faceted topics, how-tos, architecture opinions, grounded second opinions.

Both tools:

• Ground answers with Google Search (never answer from memory alone).
• Can read specific URLs you pass via the Gemini url_context tool — include background, constraints, prior conclusions, docs URLs in the query; more context = better answer.
• Resolve Gemini's vertexaisearch.cloud.google.com redirect links to their pure destination URLs (via the 302 location header) so only clean links enter your context.

API key

The extension needs a Gemini API key. Resolution order:

1. GEMINI_API_KEY environment variable
2. ~/.pi/agent/auth.json with a gemini entry:

   { "gemini": { "key": "..." } }
   

Get a key from Google AI Studio.

Install

pi install npm:pi-gemini-search

Or from GitHub:

pi install git:github.com/nilskluewer/pi-gemini-search
# or pin a version:
pi install git:github.com/nilskluewer/pi-gemini-search@v0.1.0

Try without installing:

pi -e git:github.com/nilskluewer/pi-gemini-search

How it differs from just using curl

You already have bash + curl. This package is for when you don't have a specific URL, or want a synthesized answer rather than a verbatim page dump:

• Quick fact / no URL / broad research → web_search / web_research (synthesized, lean, lossy).
• Specific doc page you need verbatim → fetch the .md with curl (full fidelity). Many Mintlify-hosted docs serve clean Markdown when you append .md to the URL.

License

MIT