Maige

You write instructions the way you'd brief a new teammate, then Maige turns each one into an action on GitHub events. A rule like labelling environment file changes as needing approval runs on every pull request that matches, with no YAML to learn and no workflow file to maintain. These natural language workflows replace the hand-sorting that fills a maintainer's morning. Rules can be edited from the dashboard when something misfires.

When an issue opens, the agent decides on labels and an assignee from its reading of your rules and the content of the report. That covers the most tedious part of GitHub issue triage: deciding who should look at this and what bucket it belongs in. It works with the labels you already use, so existing triage habits survive.

Maige reviews incoming pull requests against instructions you supply, usually a contributing guide or a review checklist. Comments land where you'd expect from a human reviewer. The review is a first pass, not a replacement for a senior maintainer, and code that ships still passes through your normal approval and branch protection.

Every connected repository gets its own isolated sandbox. The agent clones the repo there, installs dependencies, and can run commands the way a junior developer would, whether that means running tests or trying to reproduce a reported error. Writing and committing code to a new branch then becomes possible, with the pull request attributed to the bot identity.

On install, Maige indexes the whole repository into a vector store, so questions about the project are answered from the code itself rather than from guesswork. Asking which packages a repo uses or where a config value is set returns an answer grounded in files that actually exist.

The dashboard lists what the agent did on each issue and pull request, which matters when a rule behaves differently than you expected. You can trace a bad label back to the sentence that caused it, fix the wording, and watch the next run.