MCP Local LLM Server - Context Compressor

A Model Context Protocol (MCP) server that uses Ollama as a context preprocessor to reduce token consumption in Cursor by up to 80%. This server processes large files, error logs, and codebase searches locally before sending optimized summaries to Cursor.

Features

• 🤖 Ollama Integration: Uses Ollama for local LLM processing

• 📉 Token Reduction: Reduces token usage by processing context locally

• 🔧 Multiple Tools: Context compression, code analysis, log processing, and semantic search

• 📝 MCP Prompts: Dynamic instruction injection for Cursor

• 📦 MCP Resources: Read-only data resources (config, models, tools, prompts, stats)

• ⚙️ Configurable: Customizable model, temperature, and token limits

• 🚀 Easy Setup: Simple installation and configuration

• 📡 MCP Compatible: Works with any MCP-compatible client

• 🔒 Privacy: Processes sensitive data locally, only sends summaries to Cursor

• ✅ Full MCP Support: Implements all core MCP capabilities (Tools, Prompts, Resources)

Related MCP server: Portable MCP Toolkit

Prerequisites

1. Node.js (version 18 or higher)

2. Ollama installed and running

   • Download from: https://ollama.ai/

   • Install and start Ollama service

   • Pull at least one model: ollama pull llama3

Installation

1. Clone or download this repository

2. Install dependencies:

   npm install

Configuration

The server can be configured using environment variables in the MCP client configuration file.

Provider Selection

Set LLM_PROVIDER to choose which LLM provider to use:

# Select provider: 'ollama', 'openai', 'anthropic', or 'gemini'
export LLM_PROVIDER=ollama

Provider-Specific Configuration

Ollama (default):

export LLM_PROVIDER=ollama
export OLLAMA_URL=http://localhost:11434
export MODEL_NAME=llama3

OpenAI:

export LLM_PROVIDER=openai
export OPENAI_API_KEY=sk-your-api-key
export MODEL_NAME=gpt-3.5-turbo

Anthropic:

export LLM_PROVIDER=anthropic
export ANTHROPIC_API_KEY=sk-ant-your-api-key
export MODEL_NAME=claude-3-haiku-20240307

Gemini:

export LLM_PROVIDER=gemini
export GEMINI_API_KEY=your-api-key
export MODEL_NAME=gemini-1.5-flash

Common Settings

# Maximum tokens in response (default: 256)
export MAX_TOKENS=256

# Temperature for response generation (default: 0.7)
export TEMPERATURE=0.7

See MCP_CONFIGURATION.md for detailed configuration examples.

Usage

Starting the Server

# Start the MCP server
npm start

# Or for development with auto-restart
npm run dev

Available Tools

The server provides multiple tools for context compression and code analysis:

1. analyze_huge_file

Analyzes large files locally and returns a structured summary with architecture, global variables, entry points, and main logic. Reduces token usage by processing files locally before sending to Cursor.

Parameters:

• path (required): Path to the file to analyze

Example:

{
  "name": "analyze_huge_file",
  "arguments": {
    "path": "/path/to/large-file.js"
  }
}

Returns: JSON with architecture, global_variables, entry_points, main_logic, and original_size

2. digest_error_logs

Processes error logs locally to identify patterns, remove repetitive timestamps, and group similar errors. Returns a structured summary with probable cause and statistics.

Parameters:

• log_file_path (optional): Path to the log file

• terminal_output (optional): Direct terminal output content

Example:

{
  "name": "digest_error_logs",
  "arguments": {
    "log_file_path": "/path/to/error.log"
  }
}

Returns: JSON with probable_cause, occurrences, period, error_types, and recommendation

3. codebase_discovery

Performs semantic search in the codebase to find files and specific lines where related logic is implemented. Uses local processing to reduce token usage.

Parameters:

• query (required): Semantic query about the code (e.g., "where is payment processed?")

• root_path (optional): Root directory path to search in (default: current directory)

Example:

{
  "name": "codebase_discovery",
  "arguments": {
    "query": "where is payment processed?",
    "root_path": "/path/to/project"
  }
}

Returns: JSON with files (array of file references with line numbers), total_occurrences, and summary

4. ask_llm

Ask a question to the AI model running via Ollama and get a response.

Parameters:

• question (required): The question or prompt to send to the AI model

Example:

{
  "name": "ask_llm",
  "arguments": {
    "question": "What is the capital of France?"
  }
}

5. check_llm_status

Check if Ollama is running and accessible.

Example:

{
  "name": "check_llm_status",
  "arguments": {}
}

6. think_through

Adds an extra thinking layer by analyzing tasks, considering multiple approaches, and providing structured reasoning before execution.

Parameters:

• task (required): The task, question, or problem to think through

• context (optional): Additional context about the situation

• focus_areas (optional): Specific areas to focus on (e.g., ["security", "performance"])

• output_format (optional): Format of output - "plan", "analysis", "considerations", or "structured" (default)

Example:

{
  "name": "think_through",
  "arguments": {
    "task": "Refactor authentication to use JWT",
    "context": "Current: session-based, Node.js/Express",
    "focus_areas": ["security", "maintainability"],
    "output_format": "structured"
  }
}

Available Prompts

The server provides MCP prompts that inject dynamic instructions into Cursor:

1. mcp_tool_usage_rules: Mandatory rules for using MCP tools instead of direct actions

2. token_economy_guidelines: Guidelines for maximizing token savings

3. thinking_layer_instructions: Instructions for using the thinking layer

4. context_compression_rules: Rules for using context compression tools

5. chat_end_summary_rule: Automatically stores chat summaries using memory_store tool (can be disabled via DISABLE_CHAT_SUMMARY_RULE)

Note: The chat_end_summary_rule prompt is automatically available to all projects using this MCP server. To disable it, set the environment variable DISABLE_CHAT_SUMMARY_RULE=true in your MCP configuration.

Available Resources

The server exposes read-only resources via MCP:

1. mcp://local-llm/config: Current server configuration

2. mcp://local-llm/models: List of available Ollama models

3. mcp://local-llm/tools: List of all available tools

4. mcp://local-llm/prompts: List of all available prompts

5. mcp://local-llm/usage_stats: Usage statistics and token savings info

Example:

{
  "method": "resources/read",
  "params": {
    "uri": "mcp://local-llm/config"
  }
}

MCP Client Integration

To use this server with an MCP client (like Cursor), add it to your client configuration.

Basic Configuration (Ollama)

{
  "mcpServers": {
    "local-llm": {
      "command": "node",
      "args": ["path/to/your/mcp-local-llm/src/index.js"],
      "env": {
        "LLM_PROVIDER": "ollama",
        "OLLAMA_URL": "http://localhost:11434",
        "MODEL_NAME": "llama3"
      }
    }
  }
}

Using Different Providers

The server supports multiple LLM providers. Set LLM_PROVIDER to switch:

OpenAI:

{
  "env": {
    "LLM_PROVIDER": "openai",
    "OPENAI_API_KEY": "sk-your-key",
    "MODEL_NAME": "gpt-3.5-turbo"
  }
}

Anthropic:

{
  "env": {
    "LLM_PROVIDER": "anthropic",
    "ANTHROPIC_API_KEY": "sk-ant-your-key",
    "MODEL_NAME": "claude-3-haiku-20240307"
  }
}

Gemini:

{
  "env": {
    "LLM_PROVIDER": "gemini",
    "GEMINI_API_KEY": "your-key",
    "MODEL_NAME": "gemini-1.5-flash"
  }
}

See MCP_CONFIGURATION.md for complete configuration guide.

Troubleshooting

Common Issues

1. Connection Refused Error

   • Make sure Ollama is running: ollama serve or check if the service is running

   • Verify Ollama is accessible at http://localhost:11434

   • Check if Ollama is installed: ollama --version

2. No Models Available

   • Pull a model: ollama pull llama3

   • Check available models: ollama list

   • Recommended models: llama3, deepseek-coder, codellama, mistral

3. Timeout Errors

   • Large files may take time to process (max 15 seconds)

   • Consider using smaller models for faster responses

   • Check Ollama resource allocation

4. Tool Errors

   • Tools return generic error messages (identity hiding)

   • Check server logs for detailed error information

   • Verify file paths are correct and accessible

Testing the Server

You can test the server manually by sending MCP requests:

# Test checking Ollama status
echo '{"jsonrpc": "2.0", "id": 1, "method": "tools/call", "params": {"name": "check_llm_status", "arguments": {}}}' | node src/index.js

# Test asking a question
echo '{"jsonrpc": "2.0", "id": 2, "method": "tools/call", "params": {"name": "ask_llm", "arguments": {"question": "Hello, how are you?"}}}' | node src/index.js

# Test analyzing a file
echo '{"jsonrpc": "2.0", "id": 3, "method": "tools/call", "params": {"name": "analyze_huge_file", "arguments": {"path": "src/index.js"}}}' | node src/index.js

Development

Project Structure

mcp-local-llm/
├── src/
│   ├── index.js          # Main MCP server implementation
│   └── tools/            # Tool implementations
│       ├── AnalyzeHugeFileTool.js
│       ├── DigestErrorLogsTool.js
│       ├── CodebaseDiscoveryTool.js
│       └── ... (other tools)
├── package.json          # Dependencies and scripts
└── README.md            # This file

Adding New Tools

To add new tools:

1. Create a new tool class extending BaseTool in src/tools/

2. Implement getToolDefinition() and handle() methods

3. Add the tool to src/tools/index.js exports and ALL_TOOLS array

4. The tool will be automatically registered with the MCP server

License

MIT License - feel free to use and modify as needed.

Contributing

Contributions are welcome! Please feel free to submit issues and pull requests.