Multi-LLM MCP Server

Universal LLM router for Claude Code, Claude Desktop, and Claude.ai. Send prompts to GPT-4o, Claude, Gemini, Grok, Llama, and Mistral simultaneously.

Setup

cd "/c/Users/snowb/Documents/AI tech projects/LLM-MCP"
pip install -r requirements.txt

Environment Variables

Set the API keys for every provider you want to use. Providers without a key are skipped gracefully.

| Variable | Provider | |---|---| | OPENAI_API_KEY | OpenAI (GPT-4o) | | ANTHROPIC_API_KEY | Anthropic (Claude Sonnet 4.6) | | GOOGLE_API_KEY | Google (Gemini 2.0 Flash) | | XAI_API_KEY | xAI (Grok 4) | | TOGETHER_API_KEY | Together (Llama 3.3 70B) | | MISTRAL_API_KEY | Mistral (Mistral Large) |

Register with Claude Code

claude mcp add llm-arena -- python "/c/Users/snowb/Documents/AI tech projects/LLM-MCP/server.py"

Tools

llm_broadcast

Send a prompt to all configured LLM providers simultaneously and return all responses for comparison.

• message (str, required) -- the prompt to send
• system_prompt (str, optional) -- system-level instruction
• temperature (float, default 0.7) -- sampling temperature
• providers (list[str], optional) -- subset of providers to target; omit to send to all

llm_send

Send a prompt to a single specific LLM provider.

• provider (str, required) -- one of: openai, anthropic, gemini, xai, together, mistral
• message (str, required)
• system_prompt (str, optional)
• temperature (float, default 0.7)

llm_models

List all available LLM providers and their configuration status. No parameters.

llm_cage_match

Run an automated behavioral analysis comparing how different LLMs respond to the same prompt. Detects patterns like praise loops, engagement menus, therapist questions, performative insight, and identity flattery.

• message (str, required) -- the prompt to analyze
• detect_patterns (bool, default true) -- run behavioral pattern detection on each response

Example Usage

From Claude Code:

> Use llm_broadcast to ask all models "Explain quantum entanglement in one paragraph"
> Use llm_cage_match to compare how models respond to "What makes a good leader?"
> Use llm_send to ask gemini "Translate 'hello world' to Japanese"
> Use llm_models to see which providers are configured