# Connecting to Your MCP Server on Hugging Face Spaces

This guide explains how to connect to your Gradio MCP server deployed on Hugging Face Spaces, based on real-world testing and troubleshooting.

**⚠️ Important:** This guide reflects actual testing results and working solutions, not just theoretical approaches.

---

## 1. Find Your MCP Endpoint

After deploying your Gradio app to Hugging Face Spaces, your MCP endpoint will be:

```
https://YOUR_USERNAME-YOUR_SPACE_NAME.hf.space/gradio_api/mcp/sse
```

**Example:**  
If your username is `freemansel` and your space is `mcp-sentiment`, your endpoint is:

```
https://freemansel-mcp-sentiment.hf.space/gradio_api/mcp/sse
```

**Testing the endpoint:** You can verify the SSE stream is working by visiting the URL in a browser - you should see periodic ping messages every 15 seconds.

---

## 2. ✅ Working Solutions (Tested & Verified)

### A. smolagents MCPClient ⭐ **RECOMMENDED**

**Performance:** ~0.11 seconds per request (fastest!)  
**Status:** ✅ Works perfectly  
**Protocol:** Native MCP via smolagents

Based on the [Hugging Face MCP Course](https://huggingface.co/learn/mcp-course/unit2/gradio-client), this is the proper way to use MCP.

#### Installation
```bash
pdm add "smolagents[mcp]"
# or
pip install "smolagents[mcp]"
```

#### Python Code
```python
#!/usr/bin/env python3
"""
Working MCP client using smolagents.
Performance: ~0.11 seconds per request
"""

import time
from smolagents.mcp_client import MCPClient

def analyze_sentiment_mcp(text):
    """Analyze sentiment using MCP protocol."""
    mcp_client = None
    try:
        mcp_client = MCPClient(
            {"url": "https://freemansel-mcp-sentiment.hf.space/gradio_api/mcp/sse"}
        )
        
        tools = mcp_client.get_tools()
        sentiment_tool = tools[0]  # The sentiment analysis tool
        
        result = sentiment_tool(text=text)
        return result
        
    finally:
        if mcp_client:
            mcp_client.disconnect()

# Example usage
result = analyze_sentiment_mcp("I love this product!")
print(f"Polarity: {result['polarity']}")
print(f"Assessment: {result['assessment']}")
# Output: Polarity: 0.69, Assessment: positive
```

**Why this works:** smolagents handles all the MCP protocol complexity automatically, including the SSE connection and handshake.

---

### B. Gradio Client 🔄 **BACKUP SOLUTION**

**Performance:** ~1.3 seconds per request  
**Status:** ✅ Very reliable backup  
**Protocol:** Direct Gradio API access

This bypasses MCP entirely and connects directly to the Gradio API.

#### Installation
```bash
pdm add gradio_client
# or  
pip install gradio_client
```

#### Python Code
```python
#!/usr/bin/env python3
"""
Backup solution using Gradio client.
Performance: ~1.3 seconds per request
"""

import time
from gradio_client import Client

def analyze_sentiment_gradio(text):
    """Analyze sentiment using Gradio client."""
    client = Client("https://freemansel-mcp-sentiment.hf.space")
    result = client.predict(text, api_name="/predict")
    return result

# Example usage
result = analyze_sentiment_gradio("I love this product!")
print(f"Polarity: {result['polarity']}")
print(f"Assessment: {result['assessment']}")
# Output: Polarity: 0.69, Assessment: positive
```

**Why this works:** Direct API access to Gradio's built-in endpoints, no MCP protocol involved.

---

## 3. ❌ Known Issues & Failed Approaches

### A. Low-Level MCP Clients (Don't Use)

**❌ Problem:** The following approach causes `session.initialize()` to hang indefinitely:

```python
# ❌ THIS HANGS - DON'T USE
import asyncio
from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client

async def broken_mcp_approach():
    server_params = StdioServerParameters(
        command="npx",
        args=["mcp-remote", "https://freemansel-mcp-sentiment.hf.space/gradio_api/mcp/sse"]
    )
    
    async with stdio_client(server_params) as (read, write):
        async with ClientSession(read, write) as session:
            # ❌ This line hangs forever (tested timeout: 45+ seconds)
            await session.initialize()  
```

**Root Cause:** 
- The low-level MCP client requires manual protocol handling
- `session.initialize()` times out during MCP handshake
- Even though `mcp-remote` connects successfully, the protocol negotiation fails

**Evidence:** 
- STDIO client connects instantly (0.0s)
- Session creation works fine
- Failure occurs specifically in `session.initialize()`
- Error: "unhandled errors in a TaskGroup (1 sub-exception)"

---

### B. Direct HTTP Requests (Don't Work)

**❌ Problem:** MCP endpoints don't accept standard HTTP POST requests:

```python
# ❌ THIS FAILS - DON'T USE
import requests

url = "https://freemansel-mcp-sentiment.hf.space/gradio_api/call/sentiment_analysis"
payload = {"data": ["I love this product!"]}
response = requests.post(url, json=payload)
# Result: 500 Internal Server Error or 405 Method Not Allowed
```

**Why it fails:**
- MCP servers use Server-Sent Events (SSE) protocol
- Standard REST API endpoints are not exposed when `mcp_server=True`
- The `/gradio_api/mcp/sse` endpoint only accepts SSE connections

---

### C. mcp-remote with Manual Clients (Problematic)

**⚠️ Partial Success:** `mcp-remote` proxy connects but has protocol issues:

```bash
# This establishes SSE connection but MCP protocol still fails
npx mcp-remote https://freemansel-mcp-sentiment.hf.space/gradio_api/mcp/sse
```

**What happens:**
1. ✅ HTTP attempt fails (405 Method Not Allowed) - expected
2. ✅ SSE fallback succeeds - connection established  
3. ❌ MCP protocol handshake fails - timeout in client

**Evidence from testing:**
```
[8216] Using transport strategy: http-first
[8216] Received error: Error POSTing to endpoint (HTTP 405): Method Not Allowed
[8216] Recursively reconnecting for reason: falling-back-to-alternate-transport
[8216] Using transport strategy: sse-only
[8216] Connected to remote server using SSEClientTransport
[8216] Local STDIO server running
[8216] Proxy established successfully
```

---

## 4. Performance Comparison

| Method | Speed | Reliability | Protocol | Status |
|--------|-------|-------------|----------|---------|
| **smolagents MCP** | **0.11s** | ✅ Excellent | Native MCP | ⭐ Best |
| **Gradio Client** | 1.3s | ✅ Very Good | Direct API | ✅ Backup |
| Low-level MCP | ❌ Timeout | ❌ Broken | Manual MCP | ❌ Don't use |
| HTTP Requests | ❌ 405 Error | ❌ Broken | REST | ❌ Don't use |

---

## 5. Real-World Troubleshooting

### A. Import Debugging

If you have import issues, create this debugging script:

```python
#!/usr/bin/env python3
"""Debug MCP imports"""

def check_imports():
    try:
        import smolagents.mcp_client
        print("✅ smolagents available - use smolagents approach")
        return True
    except ImportError:
        print("❌ Install: pip install 'smolagents[mcp]'")
        
    try:
        import gradio_client
        print("✅ gradio_client available - use backup approach")
        return True
    except ImportError:
        print("❌ Install: pip install gradio_client")
        
    return False

check_imports()
```

### B. Connection Testing

Test the SSE endpoint directly in your browser:
```
https://freemansel-mcp-sentiment.hf.space/gradio_api/mcp/sse
```

You should see:
```
: ping - 2025-05-23 18:36:12.808434+00:00
: ping - 2025-05-23 18:36:27.809481+00:00
```

If you don't see pings, the Hugging Face Space might be sleeping.

### C. Common Error Messages

**"unhandled errors in a TaskGroup"** → Use smolagents instead of low-level MCP  
**"HTTP 405 Method Not Allowed"** → Use Gradio client instead of HTTP requests  
**"Connection timeout"** → Check if HF Space is running, try Gradio client  
**"Import error: smolagents"** → Install with `pip install "smolagents[mcp]"`

---

## 6. Ready-to-Use Scripts

Complete, tested scripts are available in the `usage/` folder:

```bash
# Primary solution (fastest)
pdm run python usage/sentiment_mcp.py

# Backup solution (reliable)
pdm run python usage/sentiment_gradio.py

# Debug imports if needed
pdm run python usage/debug_imports.py
```

---

## 7. Cursor AI Integration

For Cursor AI, use the working smolagents approach:

```json
{
  "mcpServers": {
    "sentiment-analysis": {
      "command": "python",
      "args": [
        "-c",
        "from smolagents.mcp_client import MCPClient; client = MCPClient({'url': 'https://freemansel-mcp-sentiment.hf.space/gradio_api/mcp/sse'}); tools = client.get_tools(); result = tools[0](text=input('Text: ')); print(f'Sentiment: {result}')"
      ]
    }
  }
}
```

**Note:** This is a simplified example. For production use, create a proper MCP server script.

---

## 8. Key Learnings from Real Testing

1. **Always use high-level clients** - `smolagents.MCPClient` works, low-level `mcp.ClientSession` doesn't
2. **Follow official documentation** - The [Hugging Face MCP Course](https://huggingface.co/learn/mcp-course/unit2/gradio-client) shows the right approach
3. **Have a backup plan** - Gradio client provides reliable fallback when MCP has issues
4. **Test everything** - Theoretical solutions often don't work in practice
5. **Performance varies dramatically** - smolagents (0.11s) vs Gradio (1.3s) vs broken approaches (timeout)

---

## 9. References

- [Hugging Face MCP Course](https://huggingface.co/learn/mcp-course/unit2/gradio-client) - **Use this approach**
- [smolagents Documentation](https://huggingface.co/docs/smolagents) - High-level MCP client
- [Gradio Client Documentation](https://gradio.app/guides/getting-started-with-the-python-client/) - Backup approach
- [MCP Protocol Documentation](https://modelcontextprotocol.io/) - Low-level details

---

**✅ This guide is based on extensive real-world testing and provides only verified, working solutions!**