The breadth is the draw: transcription, video analysis and image editing usually mean three servers and three configs. The preference layer is the part to judge for yourself — it rewrites your prompts before they reach Gemini, which helps when the context guess is right and is worth watching when it is not. Verbatim transcription is the standout: it keeps the fillers most transcription tools quietly delete.
An MCP server over Google's Gemini API offering seven tools, wrapped in a preference layer that detects the context of a request, enhances the prompt accordingly, and stores what worked for next time.
- `generate_image` — images from a text prompt, with optional `temperature` and a `style` chosen from realistic, artistic, minimalistic, vibrant, photographic or abstract
- `gemini-edit-image` — edit an existing image from a natural-language instruction
- `gemini-chat` — conversation with context-aware responses
- `gemini-transcribe-audio` — MP3, WAV, FLAC, AAC, OGG, WEBM or M4A to text, with a verbatim mode that keeps fillers, repeated words and expressions such as [laughs], and `preserve_spelled_acronyms` to keep U-R-L as spoken
- `gemini-code-execute` — run Python in a sandbox
- `gemini-analyze-video` — summary, transcript, objects or a detailed pass over MP4, MOV, AVI, WEBM, MKV or FLV
- `gemini-analyze-image` — objects, text and descriptions from JPEG, PNG, WebP, HEIC, HEIF, BMP or GIF
- Context types the server recognises: `consciousness`, `code`, `debugging`, `general`, and `verbatim` for transcription
A Google Gemini API key in `GEMINI_API_KEY`. Clone the repo, `npm install`, then register `node gemini-server.js` with the key in the env block; `OUTPUT_DIR` and `DEBUG` are optional. Learned preferences are stored at `./data/tool-preferences.json`. File size limits are 20MB for images, 20MB for audio and 100MB for video. Package version 2.2.0. MIT.
One command plus a key — npx -y gemini-mcp-server, then supply credentials