Hyrax

About Hyrax

Hyrax positions itself against the actual consequence of AI-assisted coding: output went up, architecture did not. It connects to GitHub, maps the repository on its first run into a HYRAX.md file and a .hyrax/discovery directory that both humans and AI contributors read as shared context, then runs six specialised agents — security, correctness, maintainability, performance, architecture and operations — over that map, each judging from the same understanding of the system rather than from a diff in isolation. Findings are prioritised by impact and carry stable HYRAX-N references, so a team can approve, dismiss or request changes on each one instead of receiving an undifferentiated wall of suggestions. Approved improvements become real pull requests, built in an isolated worktree against the team's own tests and build, and merged only by a human. The loop is designed to compound: the codebase map updates as the repository changes, so each run starts with better context than the last and drift decreases over time. A separate MCP server gives Claude Code, Cursor and Copilot live codebase context from the same map, which is the piece that makes it useful to teams already deep in agentic editors rather than a competing surface. Pricing was deliberately simplified in August 2026 from a three-tier ladder to two plans on one usage meter, with the vendor stating explicitly that there are no feature walls — the free plan is the full product.

Key Features

Tags