Hyrax AI maps repositories into persistent markdown context and uses six specialized agents to generate verified, tested pull requests.
What the product does and how it is positioned
Hyrax AI functions as an automated codebase architect designed for engineering teams that write code alongside AI tools. Installed as a GitHub App, the system analyzes repository structures, dependencies, boundaries, and existing conventions, committing live context directly into repository files like HYRAX.md and the .hyrax/discovery folder. This shared map serves as a grounding baseline for both human engineers and external coding assistants.
Rather than applying broad, unvalidated changes across a codebase, Hyrax employs six domain agents covering security, correctness, maintainability, performance, architecture, and operations. Approved improvements undergo verification in an isolated worktree using the repository's native tests, type checks, and build steps. Hyrax halts before opening a pull request if any check fails, and never merges changes on its own.
Source-supported ways to use the product
Engineering teams connect Hyrax via GitHub to automatically generate and maintain shared architectural maps, pattern catalogs, and environment variable documentation within the repository.
Teams run audits across security, maintainability, and operations to identify token storage flaws, outdated patterns, and configuration issues, turning prioritized findings into verified pull requests.
Developers use the Hyrax MCP integration to feed live repository context and established conventions into AI coding tools like Cursor, Claude Code, and Copilot.
The documented workflow, where available
Connect the repository using the Hyrax GitHub App to trigger the initial scan and generate HYRAX.md alongside discovery files.
Inspect the issues, suggestions, and advisories produced across six engineering dimensions, selecting approved tasks to address.
Hyrax executes changes in an isolated worktree, validating against local tests, linters, builds, and a re-scan of the issue.
A human engineer reviews the resulting GitHub pull request and diff before approving and merging it into the repository.
Each merged pull request triggers an update to the repository map, syncing recent changes into HYRAX.md and discovery files.
Hyrax structures repository governance around persistent repository artifacts and strict automated verification gates. When connected to a repository, the platform outputs HYRAX.md and companion documents in .hyrax/discovery that detail module boundaries, data flows, pattern occurrences, and required environment variables. Custom engineering notes placed between designated boundary markers are preserved during subsequent synchronization runs.
Before any proposed fix or improvement appears as a GitHub pull request, it must pass through an automated verification pipeline. Operating inside an isolated worktree, the system baselines the repository, applies the edit adhering to local conventions, and runs the test, typecheck, and build commands defined in the project configuration. The system also requires passing lint checks, staying under a diff size threshold, undergoing independent agent review, and completing a re-scan to verify the issue is resolved. If any check fails, the task aborts without opening a pull request.
Checks to run with your own material and workflow
What was checked and when
Answers based on the source-checked product record
No. Hyrax never merges pull requests on its own; every generated pull request requires an engineer to review and approve the change.
Six specialized agents evaluate repositories across security, correctness, maintainability, performance, architecture, and operations.
If any required check fails during the verification gate, work halts immediately and Hyrax will not open a pull request.
All AI inference runs within the Hyrax AWS Bedrock account, and customer code is not used to train AI models.
Hyrax provides an MCP server that supplies live codebase context directly to tools such as Cursor, Copilot, and Claude Code.
