MCP-LLM Bridge

A TypeScript implementation that connects local LLMs (via Ollama) to Model Context Protocol (MCP) servers. This bridge allows open-source models to use the same tools and capabilities as Claude, enabling powerful local AI assistants.

Overview

This project bridges local Large Language Models with MCP servers that provide various capabilities like:

• Filesystem operations
• Brave web search
• GitHub interactions
• Google Drive & Gmail integration
• Memory/storage
• Image generation with Flux

The bridge translates between the LLM's outputs and the MCP's JSON-RPC protocol, allowing any Ollama-compatible model to use these tools just like Claude does.

Current Setup

• LLM: Using Qwen 2.5 7B (qwen2.5-coder:7b-instruct) through Ollama
• MCPs:
  • Filesystem operations (@modelcontextprotocol/server-filesystem)
  • Brave Search (@modelcontextprotocol/server-brave-search)
  • GitHub (@modelcontextprotocol/server-github)
  • Memory (@modelcontextprotocol/server-memory)
  • Flux image generation (@patruff/server-flux)
  • Gmail & Drive (@patruff/server-gmail-drive)

Architecture

• Bridge: Core component that manages tool registration and execution
• LLM Client: Handles Ollama interactions and formats tool calls
• MCP Client: Manages MCP server connections and JSON-RPC communication
• Tool Router: Routes requests to appropriate MCP based on tool type

Key Features

• Multi-MCP support with dynamic tool routing
• Structured output validation for tool calls
• Automatic tool detection from user prompts
• Robust process management for Ollama
• Detailed logging and error handling

Setup

1. Install Ollama and required model:

ollama pull qwen2.5-coder:7b-instruct

2. Install MCP servers:

npm install -g @modelcontextprotocol/server-filesystem
npm install -g @modelcontextprotocol/server-brave-search
npm install -g @modelcontextprotocol/server-github
npm install -g @modelcontextprotocol/server-memory
npm install -g @patruff/server-flux
npm install -g @patruff/server-gmail-drive

3. Configure credentials:
   • Set BRAVE_API_KEY for Brave Search
   • Set GITHUB_PERSONAL_ACCESS_TOKEN for GitHub
   • Set REPLICATE_API_TOKEN for Flux
   • Run Gmail/Drive MCP auth: node path/to/gmail-drive/index.js auth
   • For example node C:\Users\patru\AppData\Roaming\npm\node_modules@patruff\server-gmail-drive\dist\index.js auth

Configuration

The bridge is configured through bridge_config.json:

• MCP server definitions
• LLM settings (model, temperature, etc.)
• Tool permissions and paths

Example:

{
  "mcpServers": {
    "filesystem": {
      "command": "node",
      "args": ["path/to/server-filesystem/dist/index.js"],
      "allowedDirectory": "workspace/path"
    },
    // ... other MCP configurations
  },
  "llm": {
    "model": "qwen2.5-coder:7b-instruct",
    "baseUrl": "http://localhost:11434"
  }
}

Usage

1. Start the bridge:

npm run start

2. Available commands:
   • list-tools: Show available tools
   • Regular text: Send prompts to the LLM
   • quit: Exit the program

Example interactions:

> Search the web for "latest TypeScript features"
[Uses Brave Search MCP to find results]

> Create a new folder called "project-docs"
[Uses Filesystem MCP to create directory]

> Send an email to user@example.com
[Uses Gmail MCP to compose and send email]

Technical Details

Tool Detection

The bridge includes smart tool detection based on user input:

• Email operations: Detected by email addresses and keywords
• Drive operations: Detected by file/folder keywords
• Search operations: Contextually routed to appropriate search tool

Response Processing

Responses are processed through multiple stages:

1. LLM generates structured tool calls
2. Bridge validates and routes to appropriate MCP
3. MCP executes operation and returns result
4. Bridge formats response for user

Extended Capabilities

This bridge effectively brings Claude's tool capabilities to local models:

• Filesystem manipulation
• Web search and research
• Email and document management
• Code and GitHub interactions
• Image generation
• Persistent memory

All while running completely locally with open-source models.

Future Improvements

• Add support for more MCPs
• Implement parallel tool execution
• Add streaming responses
• Enhance error recovery
• Add conversation memory
• Support more Ollama models

Related Projects

This bridge integrates with the broader Claude ecosystem:

• Model Context Protocol (MCP)
• Claude Desktop Configuration
• Ollama Project
• Various MCP server implementations

The result is a powerful local AI assistant that can match many of Claude's capabilities while running entirely on your own hardware.