# Agent Tool Registry

> A versioned registry and selection API for agent tools. Instead of every tool, the model gets a small, permission-checked, ranked shortlist per request.

- Role: Creator
- Period: 2026 – now
- Type: Open source
- Status: active
- Stack: Python, FastAPI, SQLite, Local embeddings (bge-small)
- Source: https://github.com/kanephamphu/ai-agent-tools-orchestration
- URL: https://phamphutai.com/projects/tool-registry/

## The problem

Model accuracy drops once more than about 15–20 tools are visible at once. Real agent platforms have hundreds.

## The solution

Register tools once, then ask `/advise` for a shortlist per request:

```
Registry (SQLite, versioned)
  └─ Retrieval index (cosine over local embeddings)
       └─ Selection pipeline (permissions → shortcut → router → candidates)
            └─ Execution & feedback (arg validation, confirmation gate, audit, usage stats)
```

- **Versioned tools:** they are never mutated in place, and the full history is kept.
- **Permission checks first**, before any ranking.
- **Gated, audited execution**, with outcome capture that feeds back into ranking.
- **Runs fully locally:** local embedding model, no API keys.