any-to-mcp

# any-to-mcp A Node.js framework to bridge Large Language Models (LLMs) with Model Context Protocol (MCP) servers. ## Overview any-to-mcp is a framework that enables seamless integration between LLMs and MCP servers. It allows LLMs to execute tools via the MCP protocol without requiring desktop applications like Claude Desktop or Cursor. The framework handles parsing LLM responses for XML tags, executing the appropriate MCP tool calls, and managing the conversation flow. ## Features - **LLM Adapters**: Connect to different LLM providers (OpenAI, Anthropic) - **XML Response Parsing**: Parse LLM responses for MCP-related XML tags - **MCP Server Management**: Automatically handle MCP server connections - **Tool Execution**: Route tool calls to the appropriate MCP servers - **Step-by-step Processing**: Coordinate multi-step LLM-MCP interactions - **Follow-up Questions**: Support follow-up questions to users ## Installation ```bash npm install any-to-mcp ``` ## Quick Start ```typescript import { createAnyToMcp, CoordinatorState } from 'any-to-mcp'; async function main() { // Create and initialize the framework const coordinator = await createAnyToMcp({ llm: { provider: 'openai', apiKey: 'your-openai-api-key', model: 'gpt-4', }, // Optional MCP server configuration mcp: { name: 'example-server', url: 'http://localhost:3000/mcp', type: 'http', }, // Maximum iterations before forcing completion maxIterations: 10, }); // Start a conversation with a user message const result = await coordinator.start('Hello, can you help me find the weather in San Francisco?'); // Process the result based on state if (result.state === CoordinatorState.COMPLETED) { console.log('Final answer:', result.message); } else if (result.state === CoordinatorState.WAITING_FOR_USER) { console.log('Question for user:', result.message); // User can respond with: // const nextResult = await coordinator.provideUserResponse('User response'); } // Clean up when done await coordinator.close(); } main(); ``` ## Supported LLM Providers - **OpenAI**: GPT-3.5, GPT-4, and other OpenAI models - **Anthropic**: Claude models (Opus, Sonnet, etc.) - **Custom HTTP**: Support for custom HTTP endpoints ## XML Tags Supported The framework can detect and process the following XML tags in LLM responses: - `<use_mcp_tool>`: Execute a tool on an MCP server - `<access_mcp_resource>`: Access a resource from an MCP server - `<ask_followup_question>`: Ask a follow-up question to the user - `<attempt_completion>`: Finish the conversation with a result - `<fetch_instructions>`: Fetch instructions for a specific task - `<thinking>`: Internal thought process (ignored for execution) ## Advanced Usage ### Custom LLM Adapters ```typescript import { BaseLLMAdapter, LLMRequestOptions, LLMResponse, LLMResponseChunk } from 'any-to-mcp'; // Create a custom LLM adapter class CustomAdapter extends BaseLLMAdapter { async sendRequest(messages, options) { // Implement your custom request logic return { content: 'Response from custom LLM', model: 'custom-model', }; } async sendStreamingRequest(messages, callback, options) { // Implement your custom streaming logic } } // Use the custom adapter const coordinator = await createAnyToMcp({ llm: { provider: 'custom', // Custom adapter will be used }, }); ``` ### Working with MCP Servers ```typescript import { MCPManager } from 'any-to-mcp'; // Create an MCP manager const manager = new MCPManager({ configPath: './mcpserver.json', }); // Initialize and connect await manager.initialize(); await manager.connectAll(); // Execute a tool const result = await manager.executeTool({ serverName: 'example-server', toolName: 'example-tool', arguments: { param1: 'value1' }, }); // Clean up await manager.disconnectAll(); ``` ## Examples Check the `examples` directory for more usage examples: - `simple-chat.ts`: Basic chat interface using the framework ## API Reference ### Core Functions - `createAnyToMcp(config)`: Create and initialize the framework - `coordinator.start(userMessage)`: Start a conversation - `coordinator.provideUserResponse(response)`: Provide user input - `coordinator.getState()`: Get the current state - `coordinator.reset()`: Reset the conversation - `coordinator.close()`: Clean up resources ### State Management The coordinator can be in one of the following states: - `CoordinatorState.IDLE`: Ready for a new conversation - `CoordinatorState.PROCESSING`: Processing a request - `CoordinatorState.WAITING_FOR_USER`: Waiting for user input - `CoordinatorState.COMPLETED`: Conversation completed - `CoordinatorState.ERROR`: Error occurred ## License MIT ## Contributing Contributions are welcome! Please feel free to submit a Pull Request.