Aider MCP WebSocket Server

Quick Start

This is a real MCP (Model Context Protocol) server that lets you control Aider programmatically via WebSocket. It's like having Aider as an API.

# 1. Install
npm install
cp .env.example .env
# Add your OPENAI_API_KEY=sk-... to .env

# 2. Run the MCP server (recommended)
npm run mcp

# 3. Test it
python3 test-debug.py

What you get: Send natural language commands to Aider, get back created/edited files and full responses. Each client gets isolated workspaces. Works with any MCP-compatible client (Claude Desktop, custom apps, etc.).

Overview

A production-ready MCP (Model Context Protocol) wrapper that exposes Aider's functionality over WebSocket, enabling editor plugins and services to interact with Aider programmatically. This project provides two server implementations:

1. src/mcp-server.ts - Real MCP Protocol Server (Recommended)

   • Implements standard JSON-RPC 2.0 MCP protocol
   • Proper MCP methods: initialize, tools/list, tools/call
   • Smart completion detection (waits for Aider to finish)
   • Compatible with Claude Desktop and other MCP clients

2. src/index.ts - Custom WebSocket Wrapper

   • Custom protocol with hello/welcome handshake
   • Direct stdio pipe to Aider process
   • Token-based authentication

Architecture

┌─────────────────┐         JSON-RPC 2.0        ┌─────────────────┐
│   MCP Client    │ ◄─────────────────────────► │   MCP Server    │
│  (Any MCP app)  │         WebSocket           │  (mcp-server.ts)│
└─────────────────┘                             └────────┬────────┘
                                                         │
                                                         │ spawn()
                                                         ▼
                                                ┌─────────────────┐
                                                │  Aider Process  │
                                                │  (CLI instance) │
                                                └─────────────────┘

Features

• Full MCP Protocol Support: Standard MCP methods and JSON-RPC 2.0
• Multi-tenant Isolation: One Aider child process per client
• Automatic Completion Detection: Waits for Aider to complete tasks
• Isolated Workspaces: Each session gets its own workspace directory
• OpenAI API Integration: Works with your existing OpenAI API key
• Real-time Communication: WebSocket-based bidirectional communication
• Flexible Configuration: Environment variables and CLI arguments
• Structured Logging: JSON logging with Pino

Installation

# Clone and install
git clone <repo-url>
cd aider_mcp
npm install

# Configure environment
cp .env.example .env
# Edit .env and add your OpenAI API key:
# OPENAI_API_KEY=sk-...

Prerequisites

• Node.js 18+
• Aider installed and available in PATH
• OpenAI API key

Usage

Running the Servers

# Run MCP protocol server (recommended)
npm run mcp

# Run custom WebSocket wrapper
npm run dev

# Production build
npm run build && npm start

Testing

# Quick test of MCP server
python3 test-debug.py

# Full test with waiting
python3 test-mcp-wait.py

# JavaScript test client
node test-mcp-client.js

MCP Protocol Communication

Available Tools

The MCP server exposes these tools to clients:

1. aider_command - Execute natural language commands

   {
     "name": "aider_command",
     "arguments": {
       "command": "create a REST API with Flask"
     }
   }
   

2. aider_add_file - Add files to Aider's context

   {
     "name": "aider_add_file", 
     "arguments": {
       "filename": "app.py"
     }
   }
   

3. aider_run_command - Run shell commands through Aider

   {
     "name": "aider_run_command",
     "arguments": {
       "command": "npm test"
     }
   }
   

Connection Flow

// 1. Connect via WebSocket
const ws = new WebSocket('ws://localhost:8080');

// 2. Initialize MCP session
ws.send(JSON.stringify({
  "jsonrpc": "2.0",
  "method": "initialize",
  "params": {
    "protocolVersion": "0.1.0",
    "capabilities": {},
    "clientInfo": {
      "name": "your-client",
      "version": "1.0.0"
    }
  },
  "id": 1
}));

// 3. Call tools after initialization
ws.send(JSON.stringify({
  "jsonrpc": "2.0", 
  "method": "tools/call",
  "params": {
    "name": "aider_command",
    "arguments": {
      "command": "create a Python hello world script"
    }
  },
  "id": 2
}));

Configuration

Environment Variables

# Required
OPENAI_API_KEY=sk-...        # Your OpenAI API key

# Optional
PORT=8080                    # WebSocket server port (MCP server default)
PORT=7011                    # WebSocket server port (custom wrapper default)
AIDER_MODEL=gpt-4o-mini     # Default model
WORKSPACES_DIR=./workspaces # Where to create project directories
LOG_LEVEL=info              # Logging verbosity (debug|info|warn|error)
VALID_TOKENS=token1,token2  # Auth tokens for custom wrapper (empty = no auth)

CLI Arguments

Override configuration at runtime:

# Custom port and workspace
npm run mcp -- --port 9000
npm run dev -- --port 8080 --workspaces ./custom-workspaces

# Override environment variables (custom wrapper only)
npm run dev -- --set AIDER_MODEL=gpt-4 --set LOG_LEVEL=debug

Integration Examples

Claude Desktop App

Add to Claude Desktop config:

{
  "mcpServers": {
    "aider": {
      "command": "node",
      "args": ["/path/to/aider_mcp/dist/mcp-server.js"],
      "env": {
        "OPENAI_API_KEY": "your-key",
        "AIDER_MODEL": "gpt-4o-mini"
      }
    }
  }
}

Custom MCP Client

import { WebSocket } from 'ws';

class AiderMCPClient {
  constructor(url = 'ws://localhost:8080') {
    this.ws = new WebSocket(url);
    this.requestId = 1;
    this.setup();
  }

  async callTool(toolName, args) {
    return this.sendRequest('tools/call', {
      name: toolName,
      arguments: args
    });
  }

  sendRequest(method, params) {
    const id = this.requestId++;
    this.ws.send(JSON.stringify({
      jsonrpc: '2.0',
      method,
      params,
      id
    }));
  }
}

Project Structure

aider_mcp/
├── src/
│   ├── index.ts              # Custom WebSocket wrapper
│   └── mcp-server.ts         # Real MCP protocol server
├── workspaces/               # Auto-generated client workspaces  
├── test-*.py/js              # Various test clients
├── package.json              # Node.js dependencies
├── tsconfig.json             # TypeScript configuration
├── .env                      # Environment configuration
└── README.md                 # This file

Scripts

• npm run mcp - Start MCP protocol server (recommended)
• npm run dev - Start custom WebSocket wrapper with hot reload
• npm run build - Compile TypeScript to JavaScript
• npm start - Start production server
• npm run lint - Run ESLint
• npm run typecheck - Run TypeScript type checking

Troubleshooting

Common Issues

1. "ConnectionRefusedError" or "Connect call failed"

   • Server isn't running. Start it with npm run mcp
   • Wrong port. Check .env file (MCP default: 8080, Custom default: 7011)

2. "Aider not found"

   # Install Aider
   pip install aider-chat
   
   # Verify installation
   which aider
   

3. "No API key" or Aider errors

   • Add your OpenAI API key to .env:

   OPENAI_API_KEY=sk-proj-...
   

4. Server returns too quickly / incomplete responses

   • Fixed in MCP server - waits for Aider to complete (3-second silence detection)

5. "Tool execution timeout"

   • Normal for complex tasks
   • The MCP server waits as long as needed

Debug Mode

# Run with debug logging
LOG_LEVEL=debug npm run mcp

Checking Results

Files created by Aider are stored in workspace directories:

# List all workspaces
ls -la workspaces/

# Check latest workspace  
ls -la workspaces/mcp-*/

# View created files
cat workspaces/mcp-*/your-file.py

Security Considerations

• Isolated Workspaces: Each client session gets its own directory
• Authentication: Custom wrapper supports token-based auth via VALID_TOKENS
• Git Disabled: Runs with --no-git by default for safety
• API Key Security: Never commit .env file
• Production: Consider running in Docker for additional isolation

Performance Notes

• First request to Aider may take longer (model loading)
• Subsequent requests are faster
• The server maintains one Aider process per session
• Workspace cleanup is manual (delete old directories as needed)

License

MIT