External Models MCP

A Model Context Protocol (MCP) server that empowers AI agents with a "second brain". This extension enables the Gemini CLI (and other MCP clients) to seamlessly delegate coding, reasoning, and brainstorming tasks to external state-of-the-art models like Kimi K2 (Moonshot AI), GLM-4 (Zhipu AI), and Minimax.

It acts as a local bridge, allowing your primary agent to consult specialized models for specific tasks, effectively creating a "squad" of expert AI developers.

Features

• Multi-Provider Support: Seamlessly connects to Groq, OpenRouter, and Hugging Face Inference API via a unified interface.
• Smart Context Injection: Automatically reads local project files, validates them for safety (binary detection, size limits), and formats them into XML structures that coding models understand best.
• Role-Based Routing: Define specific aliases (e.g., 'kimi' for coding, 'glm' for reasoning) to route tasks to the best model for the job.
• Expert Comparison: The compare_experts tool allows you to run the same prompt against multiple models simultaneously to verify solutions and get diverse perspectives.
• Security First: Runs entirely locally. API keys are managed via .env and are never exposed in the terminal command history.

Installation

You can install and run this server directly using uv.

1. Install via uv tool (Recommended)

uv tool install .

(Run this command from inside the project directory)

2. Configure API Keys

Create a .env file in the directory where you will run the server (or set these as system environment variables):

OPENROUTER_API_KEY=sk-or-...
GROQ_API_KEY=gsk_...
HUGGINGFACE_API_KEY=hf_...

Usage with Gemini CLI

Once installed, you can register the server in your Gemini CLI configuration using the exposed command external-models-server.

gemini mcp add external-brain --command "external-models-server"

Development / Manual Run

If you are modifying the code, you can run it directly from the source:

uv run external_models_mcp.server

Configuration

The server is pre-configured with aliases for high-performance coding models. You can modify src/external_models_mcp/server.py to add or change aliases.

Default Aliases:

• kimi-k2 (Groq): Moonshot AI's Kimi K2 (Optimized for coding)
• glm (OpenRouter): Zhipu AI's GLM-4 Flash
• minimax (OpenRouter): Minimax abab6.5s
• hf-glm (Hugging Face): GLM-4-9b-Chat via Inference API

Example Prompts

Ask a specific expert: "Use ask_expert with model='kimi-k2' to refactor src/main.py for better error handling."

Compare solutions: "Use compare_experts with experts=['kimi-k2', 'hf-glm'] to write a Python function that implements A* search algorithm."

License

MIT