Jira story in.
Evidence-backed PR out.
PatchForge is a local-first AI engineer that drives your own Claude Code install. It implements stories in isolated worktrees, proves every acceptance criterion with screenshots and video, and never publishes a thing until you approve.
How it works
Two commands. Full receipts.
PatchForge runs the whole pipeline on your machine and stops at a hard approval gate — you stay in charge of everything that ships.
Label a Jira story
Assign yourself a story and add the opt-in label. Only stories you've explicitly marked are ever eligible. That's the whole trigger.
Preflight checks everything
A fresh git worktree, dependencies installed, dev server booted, baseline tests green. If anything is broken, the run stops before a single token is spent.
Claude Code implements
Your own Claude Code install works in the isolated worktree — incremental commits, tests run and fixed. Your source never leaves your machine.
Every criterion gets proven
Playwright walks the running app and verifies each acceptance criterion — a screenshot per criterion, plus a video of the full walkthrough.
Stops at ready-for-review
The pipeline never pushes branches or opens PRs on its own. It parks the run with all its evidence and waits for you.
Local-first by design
No backend. No database. No leaks.
Every run is a folder on your disk you can inspect with ls and cat. The only network calls are Jira, GitHub, and your own Claude login.
Features
Everything your AI engineer needs
From picking up a story to delivering a reviewable, evidence-backed PR — without your code ever leaving your machine.
Human Approval Gate
The pipeline never pushes branches or opens PRs on its own. Every run parks at ready-for-review with its evidence, and only your explicit approve publishes anything.
No API Key, Ever
PatchForge drives the Claude Code CLI you already have installed, on your existing subscription. No Anthropic API key, no model credentials, no separate bill.
Evidence or It Didn't Happen
Playwright verifies each acceptance criterion in the running app — one screenshot per criterion plus a video walkthrough, all embedded in the PR body on approve.
Fail-Fast Preflight
Fresh worktree, deps installed, dev server booted, baseline tests green — verified before the agent starts. Broken environments stop the run with a clear reason.
Multi-Repo Coordination
Stories spanning frontend, backend, and shared libraries produce one branch and one PR per repo, cross-linked, with the merge order stated explicitly.
Revise Loop
Not happy with a run? Send plain-language feedback and PatchForge does another pass in the same worktree, re-validates, and brings back fresh evidence.
Inspectable State
No database, no hosted service. A run is a folder: run.json, the raw agent stream, screenshots, and video. If you can read a filesystem, you can audit PatchForge.
Watch Mode
Leave PatchForge watching your Jira queue. New labeled stories kick off runs automatically and you get notified the moment evidence is ready to review.