What is Parry? A sub-1ms pre-commit firewall for AI-written code

Hey everyone,

Parry is a deterministic static analyzer and pre-commit guardrail built for code written with Cursor, Claude Code, and Copilot. It runs locally in under 1ms with zero LLM calls and no API keys.

Why standard linters miss AI bugs

Tools like ESLint, Ruff, and cargo check were built to catch human typos and style slips. AI coding agents make completely different mistakes, and their output usually compiles with zero warnings:

- Silent permission and logic downgrades: To clear a type or trait error, an LLM might strip an auth check or swap a verified Rust Anchor Signer for a raw AccountInfo. It compiles cleanly while leaving a security hole.

- Phantom packages (slopsquatting): LLMs invent nonexistent npm, PyPI, or Rust package names. Attackers claim those names to execute malicious install scripts on your machine.

- Silent truncations: During large refactors, models sometimes drop error handlers or leave stubbed function bodies that pass basic syntax checks.

- Unbounded AI agent actions: Wiring unvalidated prompt outputs straight into database writes or on-chain transaction signers without deterministic bounds.

How Parry works

Instead of paying another slow, probabilistic LLM to review your code, Parry parses your local AST in sub-millisecond time before git commit finishes. It supports Python, TypeScript, JavaScript, and Rust out of the box, and exports native SARIF for GitHub Actions and CI.

Try the free CLI or test the rules in your browser at https://parry.scopeledger.net

We are launching on Product Hunt today (Nov 12). Run it on your repo and let us know what it catches, or share the sneakiest bug an AI model ever slipped past your linter.