# Cursor IDE Integration with MCP Server
This guide helps you integrate your local LLM MCP server with Cursor IDE.
## 🎯 Integration Options
### Option 1: SSH Tunnel Integration (Recommended)
Connect Cursor to your GPU server via SSH tunnel.
### Option 2: HTTP Bridge Integration
Run an HTTP-to-MCP bridge locally.
### Option 3: Local Installation
Install the MCP server locally (CPU-only mode).
## 🔧 Setup Instructions
### Option 1: SSH Tunnel Integration
**Step 1: Configure SSH Tunnel**
```bash
# On your local machine (where Cursor runs)
ssh -L 8080:localhost:8080 -N sazzad@your-gpu-server-ip &
# Or use the tunnel script
./deploy/ssh-tunnel-setup.sh setup
```
**Step 2: Create Cursor MCP Configuration**
Cursor uses similar configuration to Claude Desktop. Create a configuration file:
**macOS:** `~/Library/Application Support/Cursor/User/settings.json`
**Windows:** `%APPDATA%\Cursor\User\settings.json`
**Linux:** `~/.config/Cursor/User/settings.json`
Add this to your Cursor settings:
```json
{
"mcp.servers": {
"local-llm": {
"command": "ssh",
"args": [
"-t",
"sazzad@your-gpu-server-ip",
"cd /home/sazzad/mcp-server && source .venv/bin/activate && python mcp_server.py"
],
"env": {
"MODEL_NAME": "Qwen/Qwen3-8B"
}
}
}
}
```
### Option 2: HTTP Bridge Integration
**Step 1: Create HTTP Bridge**
This runs locally and forwards requests to your GPU server via HTTP.
**Step 2: Install Bridge Dependencies**
```bash
pip install aiohttp aiohttp-cors requests
```
**Step 3: Configure Cursor**
```json
{
"mcp.servers": {
"local-llm": {
"command": "python",
"args": ["/path/to/mcp-http-bridge.py"],
"env": {
"GPU_SERVER_URL": "http://your-gpu-server-ip:8080"
}
}
}
}
```
### Option 3: Local CPU Installation
For testing or if you want CPU-only mode:
```json
{
"mcp.servers": {
"local-llm": {
"command": "python",
"args": ["/Users/your-username/mcp-server/mcp_server.py"],
"env": {
"MODEL_NAME": "microsoft/DialoGPT-small"
}
}
}
}
```
## 🧪 Testing Integration
1. **Restart Cursor** after adding configuration
2. **Open Command Palette** (Cmd/Ctrl + Shift + P)
3. **Look for MCP commands** or check if tools appear
4. **Test with a simple query**
## 🔍 Troubleshooting
### Common Issues
1. **SSH Connection Issues**
- Test SSH access: `ssh sazzad@your-gpu-server-ip`
- Check SSH key authentication
- Verify server is reachable
2. **Configuration Not Loading**
- Check JSON syntax in settings.json
- Restart Cursor completely
- Check Cursor logs for errors
3. **MCP Server Not Responding**
- Ensure server is running on GPU machine
- Check if model is loaded successfully
- Verify network connectivity
### Debug Commands
```bash
# Test SSH connection
ssh -v sazzad@your-gpu-server-ip
# Test MCP server directly
echo '{"jsonrpc": "2.0", "id": 1, "method": "tools/list", "params": {}}' | ssh sazzad@your-gpu-server-ip "cd mcp-server && source .venv/bin/activate && python mcp_server.py"
# Check Cursor logs (location varies by OS)
# macOS: ~/Library/Logs/Cursor/
# Windows: %APPDATA%\Cursor\logs\
# Linux: ~/.config/Cursor/logs/
```
## 🎯 Expected Behavior
Once configured correctly, you should be able to:
- ✅ Access your local LLM through Cursor's chat interface
- ✅ Use code analysis tools powered by your GPU server
- ✅ Generate text and translations
- ✅ Summarize content using your Qwen3-8B model
## 🔧 Performance Notes
- **SSH Tunnel**: Adds ~1-5ms latency per request
- **HTTP Bridge**: Slightly higher latency but more robust
- **Local CPU**: No network latency but much slower inference
Choose the option that best fits your workflow and performance needs!
