MCP Multi Model

What mcp-multi-model does

mcp-multi-model exposes configurable model providers as MCP tools for an AI coding agent or another MCP host. It supports direct queries, parallel comparisons, category-based delegation, translation, research, health checks, and media generation.

The provider set includes OpenAI, Gemini, DeepSeek, Kimi, Grok, Perplexity, OpenRouter-hosted Anthropic models, and several other OpenAI-compatible services. Local endpoints such as Ollama, LM Studio, llama.cpp, and vLLM can also be configured without an API key. Providers without credentials are skipped automatically.

How it works

Tools are generated from the server's YAML configuration. The default setup includes a unified ask_ai entry, provider-specific query tools, ask_all, ask_both, delegate, research, translate, generate_image, generate_video, and check_health.

delegate selects a configured model according to a task category such as code, research, creative work, or realtime information. ask_all and ask_both send requests in parallel so responses can be compared. Chat tools accept an effort setting, with fast and deep mapped to provider-specific reasoning controls.

Web-enabled configurations use provider features rather than a separate browser server. Gemini can use Google Search grounding, OpenAI Responses models can use web search with source URLs, and Kimi can search the Chinese web through a tool-calling loop.

Setup and configuration

The quickest installation runs the package through npx:

{
  "mcpServers": {
    "multi-model": {
      "command": "npx",
      "args": ["-y", "mcp-multi-model"],
      "env": {
        "DEEPSEEK_API_KEY": "sk-...",
        "GEMINI_API_KEY": "AI..."
      }
    }
  }
}

API credentials can be placed in the MCP host configuration or in a .env file in the project directory. A local installation is also available by cloning the repository, running npm install, and then using npm run setup for the interactive setup wizard. The server can subsequently be started with Node and its index.js file.

Provider definitions, routing categories, and model choices are configured in config.yaml; an example configuration contains more than a dozen provider options. Adding a model involves adding a configuration block.

Tools and capabilities

• Query a selected model with ask_ai or a provider-specific tool.
• Compare every configured model with ask_all, or compare two with ask_both.
• Delegate a self-contained task to a category-selected model.
• Generate images with Gemini image models and videos with Veo.
• Run web-backed research using configured search-capable providers.
• Translate between Chinese and English.
• Check provider availability and latency with check_health.
• Control supported reasoning behavior with effort: "fast" or effort: "deep".

Limitations and notes

The server communicates over stdio MCP, so the host controls execution timeouts. Parallel queries can take several seconds, deeper requests can take longer, and video generation may take several minutes. Some clients also differ in whether they read the server's built-in instructions.

Provider quality, speed, pricing, search behavior, and supported parameters vary by model. The adapters handle some provider differences, including OpenAI reasoning parameters and unsupported temperature settings. Local providers require their own compatible endpoint to be available.

The included benchmark used a small set of tasks and samples, so its scores should be treated as configuration guidance rather than a general evaluation of every model or workload.