Summary
A second model, with its own web access, on call.
The split is by job, not by model name: search when the answer has to be current, reason when it has to be worked through, code for a repository, fileops for whatever arrives as a file. Because the file goes to Gemini directly, a large PDF or CSV never has to pass through your main model's context to be summarised.
What it is
A bridge between Google's Gemini models and MCP clients. Four tools split the work by what you are asking for rather than by which Gemini variant answers it, and the server picks the model for each call.
What you get
- gemini_search — web-connected retrieval, for questions whose answer has to be current
- gemini_reason — complex problems worked through in steps, with the steps shown
- gemini_code — full repository analysis, solution generation and debugging
- gemini_fileops — 60+ file formats including PDFs and images, several files in one call
- A model chosen per task, overridable per call or through DEFAULT_MODEL_ID
- Retries with exponential backoff and a fallback model when a call is rate-limited
Requirements
A Gemini API key as GEMINI_API_KEY, and Node. Set GEMINI_PAID_TIER when your key is on the paid tier so the higher rate limits apply. A published Docker image is available if you would rather not run it from source.