Every tool calls the same grounded model; what differs is the instruction wrapped around your input, which is why the snippet tool returns code without commentary and the comparison tool returns a table. That is genuinely useful and also the limit: the author says plainly that comparisons, architecture guidance and security findings are model output to check against the cited primary sources, not deterministic results. Two practical notes before wiring it up — a tool appearing in the list does not prove its provider is configured and reachable, and the default safety setting is BLOCK_NONE.
A research layer over a Gemini model with Google Search grounding, exposing task-shaped calls rather than one generic query box. It runs against either Vertex AI or the Gemini API, chosen at startup.
- A natural-language query answered against current search results rather than training data (answer_query_websearch).
- Documentation answered as prose or as code: a comprehensive explanation synthesised primarily from official documentation, or the exact snippet with the explanation stripped out (explain_topic_with_docs, get_doc_snippets).
- A code snippet compared against documented best practice, reporting potential bugs, performance issues and security vulnerabilities with citations and severity, scoped by language and analysis focus (code_analysis_with_docs).
- Technologies compared across criteria you name, returned as tables with pros, cons and use cases, including version-specific and compatibility notes (technical_comparison).
- Architecture patterns proposed for a described use case, with implementation examples and the tradeoffs behind each one (architecture_pattern_recommendation).
- A structured guidelines document generated from a list of technologies and versions, synthesised from current official documentation and style guides into actionable rules (generate_project_guidelines).
Node.js 18 or newer and Bun. Then one of two providers, set with AI_PROVIDER: `vertex` needs a Google Cloud project with billing and the Vertex AI API enabled, GOOGLE_CLOUD_PROJECT set, and Application Default Credentials configured; `gemini` needs only GEMINI_API_KEY. GOOGLE_CLOUD_LOCATION defaults to us-central1, and the model, temperature, streaming, max output tokens and retry behaviour are all environment variables. A Docker image is the other route — with credentials mounted read-only rather than baked in. Note that safety filters are set to BLOCK_NONE by default.
One command plus a key — bunx google-ai-search-mcp, then supply credentials