Relay — A self-hosted engineer on your Claude subscription
Takes a ticket to a reviewed pull request: proves the bug, fixes it, checks it in a browser and lets verifiers try to break it.
Request an invite · See Forge today
An engineer that has to prove the bug, then survive its own verifiers.
- Proves the defect first: Before it fixes anything, Relay writes a test that fails on the base commit. The fix has to turn it green.
- Verifiers that try to break it: Independent verifiers start fresh conversations, each told to refute the fix. The pull request opens only if none can.
- Checked in a real browser: It boots your app, drives a browser and records before-and-after captures. Docs- and test-only changes are exempt.
- Asks before it destroys: Edits land on their own. Deletes, force-pushes and schema drops wait for a person, and every decision is audited.
- Watch it live, steer it mid-run: One session shows every tool call, command, screenshot and verdict as it happens. Redirect it, or ask questions afterwards.
- Stops at its budget: Every run has a cost ceiling, and a workspace budget can hard-stop all work.
From a ticket to a pull request nobody could break.
- File the work: Describe it, attach a Linear or GitHub ticket, or let a webhook or automation file it.
- Plan and prove: Relay reads the repository in an isolated worktree and writes a test that fails on the base commit.
- Fix, test and look: It makes the fix, runs your suite and captures the app in a real browser.
- Survive, then open the PR: Verifiers try to refute the change. Only a fix that survives becomes a pull request, which Relay then reviews and keeps in sync.
Details
- Status
- v2.0 released — self-hosted, invite only
- Engine
- Claude Agent SDK, on your Claude subscription — no API key anywhere
- Starts from
- Linear or GitHub tickets, webhooks, automations, or a description
- Integrations
- GitHub App, Linear, Slack, Discord, SMTP email
- Model modes
- Auto, Max, Balanced or Lite — per issue or per workspace
- Privacy
- Issue text anonymised before the model sees it, and it fails closed
- Stack
- Django, Celery, PostgreSQL, Redis; React and TypeScript
More from ShipIt
- Vouchline — Research agents that cite and re-check every claim.
- Forge — GitHub coding agent that fixes issues and reviews PRs.
- Sentinel — Autonomous AI penetration testing with real proof.
- shipit-agent — The Python runtime under every ShipIt agent.
- Watchtower — Tracing, cost and PII masking for LLM apps.
- ShipIt Cloud — Multi-tenant console for white-label agents.
- ShipIt CLI — Terminal coding agent with two-way chat bots.
- Pentest skills — One /pentest command for Claude Code.
- ShipitQA — Goal-driven QA in real browsers, with evidence.
- KeyVault — Zero-knowledge password manager. No server.
- MacSweep — Clean, protect and understand your Mac.
- Snappilot — Free screen capture studio, fully on-device.
- shipit-ui-design — Senior UI/UX as a Claude Code plugin.
- ShipIt Palette — One hex to a full design system.