llm-mcp

A Ruby gem that exposes Large Language Models (LLMs) via the Model Context Protocol (MCP), enabling seamless integration of AI capabilities into your development workflow.

llm-mcp creates an MCP server that provides standardized access to various LLM providers (OpenAI, Google Gemini, and OpenAI-compatible APIs) while supporting advanced features like session management, conversation persistence, and integration with external MCP tools.

- 🤖Multi-Provider Support: Works with OpenAI, Google Gemini, and any OpenAI-compatible API
- 💬Session Management: Persist conversations across server restarts
- 🔧MCP Tool Integration: Connect to external MCP servers and use their tools within LLM conversations
- 📝Comprehensive Logging: JSON-formatted logs for debugging and analysis
- 🔌Extensible Architecture: Easy to add new providers and customize behavior
- 🚀Built on FastMCP: Leverages the fast and efficient MCP server framework

Add this line to your application's Gemfile:

Set up your API keys based on the provider you want to use:

# For OpenAI export OPENAI_API_KEY="your-openai-api-key" # For Google Gemini export GEMINI_API_KEY="your-gemini-api-key" # or export GOOGLE_API_KEY="your-google-api-key"

Start an MCP server that exposes an LLM:

# Using OpenAI llm-mcp mcp-serve --provider openai --model gpt-4 # Using Google Gemini llm-mcp mcp-serve --provider google --model gemini-1.5-flash # Using a custom OpenAI-compatible API llm-mcp mcp-serve --provider openai --model llama-3.1-8b --base-url https://api.groq.com/openai/v1

llm-mcp mcp-serve \ --provider openai \ --model gpt-4 \ --verbose \ # Enable verbose logging --json-log-path logs/llm.json \ # Log to JSON file --session-id my-project \ # Resume a specific session --session-path ~/my-sessions \ # Custom session storage location --append-system-prompt "You are a Ruby expert" \ # Add to system prompt --skip-model-validation # Skip model name validation

llm-mcp can connect to other MCP servers, allowing the LLM to use their tools:
- Create an MCP configuration file (e.g.,~/.mcp/config.json):

{ "mcpServers": { "filesystem": { "command": "npx", "args": ["@modelcontextprotocol/server-filesystem", "/tmp"] }, "github": { "command": "mcp-github", "env": { "GITHUB_TOKEN": "your-github-token" } }, "http-api": { "url": "https://api.example.com/mcp/sse", "transport": "sse", "headers": { "Authorization": "Bearer your-token" } } } }

llm-mcp mcp-serve \ --provider openai \ --model gpt-4 \ --mcp-config ~/.mcp/config.json

Now the LLM can use tools from the connected MCP servers in its responses!

Send a request to the LLM and get a response.

- prompt(required): The message or question for the LLM
- temperature(optional): Control randomness (0.0-2.0, default: 0.7)
- max_tokens(optional): Maximum response length

{ "method": "tools/call", "params": { "name": "task", "arguments": { "prompt": "Explain the concept of dependency injection", "temperature": 0.7, "max_tokens": 500 } } }

Clear the conversation history and start fresh.

{ "method": "tools/call", "params": { "name": "reset_session", "arguments": {} } }

Sessions automatically persist conversations to disk, allowing you to:

- Resume previous conversations
- Maintain context across server restarts
- Track token usage over time

Sessions are stored in~/.llm-mcp/sessions/by default, with each session saved as a JSON file.

- Message history (user, assistant, and system messages)
- Timestamps for each interaction
- Token usage statistics
- Session metadata

Enable JSON logging for comprehensive debugging:

llm-mcp mcp-serve \ --provider openai \ --model gpt-4 \ --json-log-path logs/llm.json \ --verbose

- All requests and responses
- Tool calls and their results
- Session operations
- Error messages and stack traces

Add to your Claude Desktop configuration (~/Library/Application Support/Claude/claude_desktop_config.json):

{ "mcpServers": { "llm-mcp": { "command": "llm-mcp", "args": ["mcp-serve", "--provider", "openai", "--model", "gpt-4"], "env": { "OPENAI_API_KEY": "your-api-key" } } } }

require 'mcp-client' client = MCP::Client.new client.connect_stdio('llm-mcp', 'mcp-serve', '--provider', 'openai', '--model', 'gpt-4') # Use the task tool response = client.call_tool('task', { prompt: "Write a haiku about Ruby programming", temperature: 0.9 }) puts response.content

Create a powerful AI assistant by combining llm-mcp with other MCP servers:

{ "mcpServers": { "llm": { "command": "llm-mcp", "args": ["mcp-serve", "--provider", "openai", "--model", "gpt-4", "--mcp-config", "mcp-tools.json"] }, "filesystem": { "command": "mcp-filesystem", "args": ["/project"] }, "git": { "command": "mcp-git" } } }

After checking out the repo, runbin/setupto install dependencies. Then, runrake testto run the tests.

# Install dependencies bundle install # Run tests bundle exec rake test # Run linter bundle exec rubocop -A # Install gem locally bundle exec rake install

Bug reports and pull requests are welcome on GitHub athttps://github.com/parruda/llm-mcp.

Anchor Browser (https://anchorbrowser.io) is secure infrastructure for computer-use agents — stealth cloud browsers, authentication, captcha bypass, and a hosted MCP server for Cursor, Claude, and Windsurf.