# MCP Agent Reliability Server

**Daily AI Project by Grok for Prince Ruhul / Prevalid**  
Date: 2026-07-29

## What problem does this solve?

AI agents call tools all day. Sometimes they succeed, sometimes they fail silently.  
Entrepreneurs and founders have no easy way to see **"Is my agent actually reliable?"**

This MCP server gives any AI agent (Claude, Cursor, Gemini, etc.) simple tools to:

- Log every tool call result (success / fail + latency)
- Calculate a live reliability score (0–100)
- Generate a clear reliability report
- Detect patterns of failure

All pure computation — no paid API required. Runs locally or on any serverless.

## Quick Start

```bash
# Install
pip install -e .

# Run the MCP server (stdio)
python -m src.server

# Or with uv
uvx mcp-agent-reliability
```

Add to your Claude Desktop / Cursor MCP config:

```json
{
  "mcpServers": {
    "agent-reliability": {
      "command": "python",
      "args": ["-m", "src.server"],
      "cwd": "/path/to/this/repo"
    }
  }
}
```

## Tools exposed

| Tool | Description |
|------|-------------|
| `log_tool_call` | Record a tool call outcome (name, success, latency_ms, optional error) |
| `get_reliability_score` | Get overall or per-tool reliability score (0-100) |
| `generate_reliability_report` | Full markdown report with scores, top failures, recommendations |
| `reset_stats` | Clear all logged data (for testing) |
| `list_tracked_tools` | See which tools have been logged so far |

## Why this is valuable for entrepreneurs

- You can finally **measure** if your AI agent is production-ready
- Agents can self-report their health before critical tasks
- Zero cost — no external API keys needed
- Works with the new **stateless MCP** (July 28 2026 update)

## License

MIT

Built as part of the Daily AI Project series for Prevalid.