Pi Safety

A library that classifies AI agent tool calls as READ (no side effects) or WRITE (everything else) using glob-based command pattern matching.

When an AI agent runs npx go-gmail marc@example.com search "invoice", that’s a read. When it runs npx go-gmail marc@example.com send --to bob@example.com, that’s a write. Pi Safety provides the registry where skills declare their safety levels, and downstream tools (like Pi Planner) enforce them.

Patterns use glob wildcards and are registered per tool/CLI. The LLM itself acts as the parser — when it loads a skill, it extracts the safety classifications and registers them. This means any new skill automatically gets safety enforcement without manual configuration.

Part of the pi-mf-extensions monorepo.