The value is not just reaching Gemini — it is that temperature, system instruction, safety settings, JSON schema and grounding are arguments you set per call, so a structured extraction and a creative draft can come from the same server without reconfiguring anything. Two habits pay off: check list_models rather than hard-coding a name, since the model is an argument and the roster moves, and run count_tokens before pushing a large document at a long-context model.
A stdio bridge to the Google Gemini API. Six tools cover generation, vision, embeddings, token counting, model discovery and built-in help, with the model named per call.
- generate_text with the parameters exposed as arguments: model, temperature, topK, topP, maxTokens, systemInstruction and safetySettings
- jsonMode with a jsonSchema, so a call can return a structured object instead of prose you then have to parse
- Grounding turns on Google Search grounding for questions where current material matters
- conversationId keeps a thread going across calls rather than restarting the context each time
- analyze_image takes a prompt plus an image URL or base64 data
- embed_text produces embeddings from Gemini's embedding models
- count_tokens measures text against a specific model before you send it
- list_models returns the available models and their capabilities, with an optional filter
- get_help answers usage questions about the server itself, by topic
A Gemini API key from Google AI Studio, passed as GEMINI_API_KEY in the client's env block — it is read from the environment only, never an argument. It runs over stdio as npx -y github:aliargun/mcp-server-gemini, so Node and network access are all that is needed locally. Usage is billed by Google against that key. MIT.
One command plus a key — npx -y github:aliargun/mcp-server-gemini, then supply credentials