Bug reports in.
Pull requests out.
pr-goblin turns a bug report from your users or testers into a proposed fix. It reads the report, writes the code in an isolated branch, and opens a pull request. You review it and merge it, or you don't.
It never merges. Every fix is a pull request for a human, and your CI is the gate.
Most bugs are small. Getting to them is not.
A typical bug costs ten minutes to fix and an afternoon to reach: reading a vague report, asking for the browser, reproducing it, finding the file. pr-goblin does the reaching, so your team starts at the diff.
Reports that arrive complete
The widget attaches the URL, browser, viewport, a screenshot and the page's recent console and network errors. No more “which browser?” threads.
A fix waiting for you
Claude Code or Codex works each ticket in its own git worktree and opens a PR. Your morning starts with reviews, not reproductions.
Nothing ships without you
No auto-merge, ever. Reports from strangers are held until a human approves them. You keep control; the goblin keeps busy.
Four small parts. Use all of them, or one.
Each part runs on infrastructure you already have. Start with only the runner and tickets you write by hand, and add the widget later.
Widget
A feedback button that takes on your app's own colours, fonts and corners. Captures context, masks private data.
Intake
One endpoint that turns a report into a ticket. Holds the tracker token so the browser never sees it.
Triage
Classifies the report, sets severity and drafts a fix plan as a comment, before any code is written.
Runner
Picks up ready tickets, runs a coding agent in an isolated worktree, and opens the pull request.
Built for the day someone tries to trick it.
A bug report from the public is text written by a stranger, and it becomes input to an agent with access to your code. pr-goblin is designed around that fact.
- Never mergesEvery change is a pull request. Your review and your CI decide.
- Severity is set by triage, not the reporterA stranger cannot mark their own report critical to jump the queue.
- Screenshots are maskedForm values are blanked, marked elements are blacked out, and reporters can cover anything else before sending.
- Signed testers skip the queueYour backend signs a tester's login; their reports are trusted and go straight to work.
Urgent now. Everything else when it suits you.
Labels decide when a ticket is worked, so a burst of minor reports never crowds out the one that matters, and your agent budget goes where you point it.
Urgent
Linear priority Urgent, or a critical label. Worked ahead of everything else.
Normal
Everything else. Picked up within about 30 seconds when the runner runs as a daemon.
Batch
A batch label holds the ticket for the overnight window, a few per cycle. Wake up to a tray of PRs.
No new platform to adopt.
pr-goblin plugs into the tracker, agent and hosting you already pay for. Use your Claude or ChatGPT subscription for solo work, or an API key for a team.
First PR before your coffee cools.
The quickest path needs no server and no settings file: the runner on GitHub Actions, working issues you label.
- Copy one workflow into
.github/workflows/. - Add two secrets: your agent credential and a repo-scoped GitHub token.
- Label an issue
goblinand wait for the pull request.
# 1. add the workflow curl -sLo .github/workflows/goblin.yml \ https://raw.githubusercontent.com/Rumi-Shamz/pr-goblin/v0/examples/goblin.yml # 2. add the secrets gh secret set CLAUDE_CODE_OAUTH_TOKEN # from `claude setup-token` gh secret set GOBLIN_GH_TOKEN # contents, issues, PRs: read/write # 3. hand the goblin a bug gh issue edit 123 --add-label goblin
Free to run yourself. Hosted when you'd rather not.
The whole loop is open source and always will be. Hosted plans are on the way for teams who want the goblin without the upkeep.
For developers happy to run a workflow or a small daemon.
- Widget, intake, triage and runner
- GitHub Issues or Linear
- Claude Code or Codex, your credentials
- Unlimited repos and reports
- Community support on GitHub
For small teams who want it working today and maintained tomorrow.
- Everything in Self-hosted
- Managed intake endpoint, no deploy
- Always-on runner with a fast urgent lane
- Dashboard of reports, runs and PRs
- Email and Slack notifications
For organisations with many repos, policies and reviewers.
- Everything in Hosted
- SSO and per-repo policies
- Usage budgets per lane
- Audit log of every agent run
- Priority support
Fair questions.
Will it push broken code to production?
No. pr-goblin cannot merge. It opens a pull request, your CI runs on it, and a person decides. A bad fix costs you one closed PR.
What stops a malicious bug report from hijacking the agent?
Reports from the public widget are labelled untrusted and no agent touches them until a human removes that label. Urgent untrusted reports get a read-only investigation at most. Only signed reports from your own testers skip the hold.
Do I need to use all four parts?
No. Many teams start with only the runner, working issues they write and label themselves. Add the widget and intake when you want reports straight from users.
What does it cost to run?
The software is free. You pay for the coding agent: either your existing Claude or ChatGPT subscription for solo use, or an API key billed per use for a team. When a subscription hits its usage cap, tickets wait for the reset instead of failing.
What about my clients' data in screenshots?
Form values are blanked by default, any element marked data-goblin-mask (or the markers Sentry, rrweb, PostHog and FullStory use) is blacked out, and reporters can cover anything else. Screenshots are stored only in a private repo unless you say otherwise, and you can turn them off entirely.
Is it ready for production?
It is pre-release. Every part is built and unit-tested, but the full loop has not yet been run end to end against a live tracker. The spec in the repo says exactly what is and isn't verified.
Give your bug backlog a goblin.
Open source, five minutes to set up, and you stay the one who merges.