# mcp-agent-reliability

Pure-computation MCP server that scores AI agent trajectories, detects silent failures, loops, and reliability metrics — zero external API cost.

[![Available on MCPize](https://img.shields.io/badge/MCPize-Ready-blue)](https://mcpize.com)

## What problem does this solve?

AI agents often fail silently: they loop on the same tool, return empty results, or give high-confidence answers without evidence. Debugging is expensive. This MCP gives you instant scores and failure reports from any agent run log you provide.

Think of it like a doctor’s check-up for your AI agent — it looks at the “X-ray” (the run log) and tells you what is healthy and what is broken, without needing another expensive doctor (LLM).

## Tools

| Tool | Description |
|------|-------------|
| `score_trajectory_tool` | 0-100 reliability score + breakdown |
| `detect_failure_modes_tool` | List of loops, empty results, high-confidence-without-evidence, etc. |
| `analyze_tool_usage_tool` | Per-tool call counts, error rates |
| `compute_success_rate_tool` | Success rate across many runs |
| `compare_trajectories_tool` | Which of two runs is more reliable |

## Quick Start (local)

```bash
# Install
pip install -e .

# Run (HTTP on port 8080)
python -m mcp_agent_reliability.server
```

Or with MCP inspector / Claude Desktop / Cursor by pointing to the HTTP endpoint.

## Example trajectory input

```json
[
  {"tool": "get_weather", "status": "ok", "result": {"temp": 28}},
  {"role": "assistant", "content": "28C today", "is_final": true}
]
```

## Why this is valuable for entrepreneurs

- Zero running cost (no paid APIs)
- Helps you ship reliable agents faster → happier users → more revenue
- Can be called by the agent itself mid-run or by your CI after tests
- Fits the “Type A” high-margin MCP pattern preferred on MCPize

## Development

```bash
pytest
```

## License

MIT

Built daily for Prince Ruhul / Prevalid by the Daily AI Project Builder.