Cross Llm MCP

What JamesANZ/cross-llm-mcp MCP server does

The JamesANZ/cross-llm-mcp MCP server presents multiple hosted LLM APIs through a single MCP connection. It supports ChatGPT, Claude, DeepSeek, Gemini, Grok, Kimi, Perplexity, Mistral, and Hugging Face Inference Router models. The Hugging Face integration uses hosted inference and accepts a Hub model repository ID; it does not run model weights locally inside Node.js.

The server can send a prompt to one named provider, invoke a provider-specific tool, or submit the same prompt to all supported providers. It also records prompt activity and exposes tools for reviewing, filtering, analyzing, deleting, or clearing those records. Preference tools allow a default model, a cost preference, and task-oriented model mappings to be configured.

How it works

An MCP client such as Cursor or Claude Desktop starts the local Node.js server and communicates with it through MCP. Each provider call uses the corresponding API credential supplied through the process environment. Provider availability therefore depends on which credentials have been configured and on the selected provider's API access.

Model selection can use tags including coding, business, reasoning, math, creative, and general. The configured preferences can favor a default model or cheaper models, while the provider-specific tools allow direct calls. The call-all-llms tool is intended for comparing responses from the available providers, although using it can generate multiple external API requests.

The JamesANZ/cross-llm-mcp MCP server is an API aggregation layer rather than a local inference runtime. For Hugging Face, requests go to the hosted Inference Router. Running Ollama, llama.cpp, or another local inference service is outside the server's documented operation.

Setup and configuration

The documented requirements are Node.js 18 or newer and npm. The package can be installed globally with npm, or the repository can be cloned, installed, and built with npm install followed by npm run build. Cursor provides a one-click installation link, and Claude Desktop can be configured with a local Node command pointing to the built build/index.js file.

Set the credentials for the providers you intend to call. The documented variables are OPENAI_API_KEY, ANTHROPIC_API_KEY, DEEPSEEK_API_KEY, GEMINI_API_KEY, XAI_API_KEY, KIMI_API_KEY, PERPLEXITY_API_KEY, MISTRAL_API_KEY, and either HF_TOKEN or HUGGINGFACE_API_KEY. Hugging Face also supports optional model and base-URL variables. Restart Claude Desktop after changing its server configuration.

Tools and capabilities

The JamesANZ/cross-llm-mcp MCP server documents these capability groups:

• Individual calls for ChatGPT, Claude, DeepSeek, Gemini, Grok, Kimi, Perplexity, Mistral, and Hugging Face.
• call-llm for selecting a provider by name and call-all-llms for sending a prompt to every provider.
• Preference management through get-user-preferences and set-user-preferences.
• Tag-based model discovery with get-models-by-tag.
• Prompt log access through history, statistics, filtered deletion, and complete clearing tools.

The repository identifies Node.js, TypeScript, the MCP SDK, Superagent, and Zod as part of its implementation. It lists macOS, Windows, and Linux as supported platforms.