minghuig/llm-mcp-server

LLM MCP Server

A simple MCP (Model Context Protocol) server that provides tools to query Claude, ChatGPT, and Gemini.

Features

• query_claude: Send messages to Claude (Anthropic API) with multi-turn conversation support
• query_chatgpt: Send messages to ChatGPT (OpenAI API) with multi-turn conversation support
• query_gemini: Send messages to Gemini (Google API) with multi-turn conversation support
• export_conversation: Export conversations to markdown files
• AI-to-AI conversation context automatically included in system prompts
• System prompt persistence across conversations

Setup

1. Create a virtual environment:

python -m venv venv
source venv/bin/activate

2. Install dependencies:

pip install -e .

3. Create a .env file with your API keys:

cp .env.example .env
# Edit .env and add your actual API keys

Usage

Run the MCP server:

python server.py

Configuration

Configure your Claude Desktop or other MCP client to connect to this server.

Example configuration for Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "llm": {
      "command": "python",
      "args": ["/absolute/path/to/server.py"]
    }
  }
}

The server will automatically load API keys from the .env file. Alternatively, you can override them by adding an env section to the config above.

Available Tools

query_chatgpt

Send a message to ChatGPT (OpenAI). Supports multi-turn conversations using server-side context management.

Parameters:

• message (required): The message to send
• model (optional): Model to use (default: "gpt-5.1-2025-11-13")
• system_prompt (optional): System prompt for context (persists throughout conversation unless overridden)
• temperature (optional): 0.0 to 2.0 (uses model default if not specified)
• max_tokens (optional): Maximum response length
• conversation_id (optional): Conversation ID from a previous call to continue the conversation

Multi-turn conversations: The tool returns a Conversation ID in the output (format: chatgpt_...). Pass this ID to conversation_id in subsequent calls to continue the conversation. The server stores conversation history in memory with automatic limits to prevent memory issues.

Conversation IDs are provider-specific and cannot be used across different providers (e.g., a ChatGPT conversation_id cannot be used with Claude).

System prompt persistence: When you provide a system_prompt in the first message of a conversation, it will be automatically reused for all subsequent messages in that conversation unless explicitly overridden.

Example workflow:

1. First call: query_chatgpt("Hello!") → Returns [Conversation ID: chatgpt_abc123def456]
2. Follow-up: query_chatgpt("How are you?", conversation_id="chatgpt_abc123def456")

query_claude

Send a message to Claude (Anthropic). Supports multi-turn conversations using server-side context management.

Parameters:

• message (required): The message to send
• model (optional): Model to use (default: "claude-sonnet-4-5-20250929")
• system_prompt (optional): System prompt for context (persists throughout conversation unless overridden)
• temperature (optional): 0.0 to 1.0 (uses model default if not specified)
• max_tokens (optional): Maximum response length (default: 4096)
• conversation_id (optional): Conversation ID from a previous call to continue the conversation

Multi-turn conversations: The tool returns a Conversation ID in the output (format: claude_...). Pass this ID to conversation_id in subsequent calls to continue the conversation. The server stores conversation history in memory with automatic limits to prevent memory issues.

Conversation IDs are provider-specific and cannot be used across different providers (e.g., a Claude conversation_id cannot be used with Gemini).

System prompt persistence: When you provide a system_prompt in the first message of a conversation, it will be automatically reused for all subsequent messages in that conversation unless explicitly overridden.

Example workflow:

1. First call: query_claude("Hello!") → Returns [Conversation ID: claude_abc123def456]
2. Follow-up: query_claude("How are you?", conversation_id="claude_abc123def456")

query_gemini

Send a message to Gemini (Google). Supports multi-turn conversations using server-side context management.

Parameters:

• message (required): The message to send
• model (optional): Model to use (default: "gemini-2.5-flash")
• system_prompt (optional): System prompt for context (persists throughout conversation unless overridden)
• temperature (optional): 0.0 to 2.0 (uses model default if not specified)
• max_tokens (optional): Maximum response length (uses model default if not specified)
• conversation_id (optional): Conversation ID from a previous call to continue the conversation

Multi-turn conversations: The tool returns a Conversation ID in the output (format: gemini_...). Pass this ID to conversation_id in subsequent calls to continue the conversation. The server stores conversation history in memory with automatic limits to prevent memory issues.

Conversation IDs are provider-specific and cannot be used across different providers (e.g., a Gemini conversation_id cannot be used with Claude).

System prompt persistence: When you provide a system_prompt in the first message of a conversation, it will be automatically reused for all subsequent messages in that conversation unless explicitly overridden.

Example workflow:

1. First call: query_gemini("Hello!") → Returns [Conversation ID: gemini_abc123def456]
2. Follow-up: query_gemini("How are you?", conversation_id="gemini_abc123def456")

export_conversation

Export a conversation to a markdown file. The conversation will be saved to the directory specified by the LLM_CONVERSATIONS_DIR environment variable.

Parameters:

• conversation_id (required): The conversation ID to export (e.g., "chatgpt_abc123def456")
• calling_llm_name (required): The name of the LLM making the calls (e.g., "Claude", "ChatGPT", "Gemini")

Output format: The exported file will be a markdown file with the conversation ID as the filename (e.g., chatgpt_abc123def456.md). The conversation is formatted with alternating messages labeled by the participant names.