For the WebUI just open translateGemmaWebUI.html in a browser and point it to your llama.cpp endpoint running TranslateGemma.

A Model Context Protocol (MCP) server that provides high-quality text translation via the local TranslateGemma LLM API. Designed for low-latency, private, and secure on-device translation with full control over the underlying model.

• Local Translation Inference: Leverages TranslateGemma (or compatible OpenAI-like API) on your machine — no external data leaves your device.
• Multi-Language Support: Translate between any language pairs using standardized BCP-47 codes (en, de-DE, fr-FR, ja-JP, etc.).
• Automatic Retries & Rate Limiting: Built-in exponential backoff and 30 req/min rate limiting for stability.
• Robust Error Handling: Graceful failures with clear, actionable error messages.
• Browser/Client Friendly: CORS-enabled HTTP endpoint for direct integration.
• LLM-Ready Output: Clean, structured responses optimized for tool-use workflows.

• Python 3.9+
• fastmcp, uvicorn, httpx, and starlette
• A local TranslateGemma-compatible server (e.g., running via ollama, vllm, or custom server) exposing an OpenAI /v1/chat/completions endpoint.

pip install fastmcp uvicorn httpx starlette

git clone https://github.com/woheller69/TranslateGemmaMCP.git
cd TranslateGemmaMCP

Ensure your TranslateGemma API is accessible at http://127.0.0.1:8080/v1/chat/completions (or configure via --api-url).

llama-server -ngl 99 -m translategemma-27b-it.Q6_K.gguf -c 8000 --no-jinja

python3 server.py --host 0.0.0.0 --port 3000

Or override the API endpoint:

python3 server.py --api-url http://192.168.1.100:8000/v1/chat/completions

The server exposes an MCP-compatible HTTP endpoint at http://localhost:3000/mcp.

Translates text using TranslateGemma with intelligent language handling.

async def translate(
    text: str,
    source_lang_code: str,
    target_lang_code: str,
    ctx: Context,
    max_retries: int = 2,
) -> str

• ✅ Translated text (clean, trimmed)
• ❌ Error message (prefixed with ❌ / ⚠️) on failure

• All translation is performed locally on your machine.
• No telemetry, external tracking, or data collection.
• CORS configured permissively (*) for local dev — restrict origins in production.

MIT License — see LICENSE for details.