proof

# Proof

Code-validated pattern intelligence from pydantic-ai's actual source code.

## What It Does

Proof extracts working patterns from real library code via analysis — not from documentation, blog posts, or human guesses. 391 patterns from pydantic-ai and pydantic, mapped to a knowledge graph of
1,251 nodes and 8,911 edges.

When you ask Proof to build an agent, it assembles code from patterns the library's own code structure validates. If a pattern fails, your feedback makes the network smarter.

## Quick Start

No API keys. No auth. No configuration needed.
Connect via Smithery CLI
npx -y @smithery/cli add proof --transport streamable-http https://mcp.aigentys.com

Then in your AI assistant: *"Run catalog() to see what's available."*

## Tools (5)

| Tool | Description |
|------|-------------|
| **catalog()** | List all libraries, pattern counts, and available data. Start here. |
| **search(query)** | Find patterns by keyword (e.g., "agent", "tool", "validator"). |
| **explain(symbol)** | Deep-dive into a class or function — its methods, dependencies, gotchas. |
| **build_agent(description)** | Generate working agent code from validated patterns. |
| **report(about, worked, details)** | Report whether generated code worked. Feedback drives confidence scores. |

## How It Works

1. **Graph Engine** loads the knowledge graph (SurrealDB, 1,251 nodes, 8,911 edges)
2. **Pattern Extractor** finds validated patterns from AST analysis of pydantic-ai source
3. **Agent Assembler** composes working code from patterns, not hallucinated guesses
4. **Confidence Scores** start at 0.5 (code-derived) and update with user feedback

## What's in the Network

- **pydantic-ai v1.77.0** — 751 classes, 369 functions, 98 docs, 33 examples
- **pydantic v2.12.4** — type system, validation, serialization patterns
- **391 extracted patterns** — reusable code structures validated by the library itself

## Roadmap

Cross-library intelligence (langchain, crewai, OpenAI) is next. Right now it's pydantic-ai + pydantic only.

Your feedback is what makes the network smart. Use `report()` after every build.