PRKrew turns a GitHub or local issue into a planned, agent-executed pipeline — analyse, review, implement, QA — run by squads of AI agents across every repo in your project, ending in one pull request per repo. No auto-merge, ever.
Every step is visible on a live board — a graph of steps wired like a circuit, each owned by a squad of agents.
From GitHub, Gitea, JIRA — or created locally, fully offline. Labels decide what PRKrew picks up.
The issue becomes a step-by-step graph, grounded in your actual code. Optionally hold it for your approval before anything runs.
Agent squads work steps in parallel across repos — review, implement, QA — with live output, tool traces and token spend on the board.
Finished work lands as pull requests, ready for human review. You run the result, you complete the merge.
Short, narrated walkthroughs of the pieces that matter.
Our first real test case of autonomous self-development: PRKrew decides what to build next, plans it, and builds it — the only human touch is the sign-off on the pull request.
A Director agent inside PRKrew reviews the product, weighs candidate features, and files the issue for the one that helps users most. No human wrote the ticket.
Self-development runs on a dedicated machine with a budget of its own — the same guardrails and hard stops PRKrew gives every run.
Everything ends in a pull request on GitHub, reviewed and merged by a person. That's the whole human role — and it's non-negotiable.
Autonomy you can actually trust in a real codebase.
Hold runs at stage boundaries — before QA, before a PR is pushed. Approve from the board, the inbox, or a push notification on your phone.
Plans to approve, halted runs, waiting gates, budget alerts — every decision lands in one place and links straight to the work.
Live USD spend per run with warnings and hard stops. A run can't blow through the ceiling you set.
Runs through the agent CLIs already on your machine, with cross-runtime failover — cloud or fully local models.
A project spans many repositories; each step targets the repos it needs and fans out one pull request per repo.
Every run is recorded — cost, tokens, duration, failures, tool calls — with per-country rate cards that turn runs into a defensible savings number.
PRKrew is in active development at KH-IT. Get in touch for a demo or early access.