open beta · static + server-rendered · free
Deploys that fixthemselves.
Connect a repo. AutoDeploy detects how to build it, runs it in a sandbox, repairs the failures it understands, validates the result — and shows you every decision it made.
free · no credit card · rules first, AI only when needed
- runningAnalyzevite · pnpm · node 22
- queuedBuildpnpm build
- queuedRecoverrule R-004 · node engine
- queuedValidatechromium · 0 errors
- queuedLiveshop.autodeploy.fyi
- adminbuild failed · escalated2m
- bloglive38s
- docsai patch · approval1m
- shopauto-fixed41s
01 · build
Most of a deploy is waiting.
So every build streams to you live: queued with fair scheduling, built in a sandbox with no network except the package registry, validated in a real browser before anyone sees it.
- ✓Analyze
- ✓Queue
- ✓Build
- ✓Validate
- ✓Live
$pnpm install --frozen-lockfilePackages: +412Progress: resolved 412, reused 380, downloaded 32$vite buildvite v7.3.6 building for production…✓ 214 modules transformed.dist/index.html 0.46 kB │ gzip: 0.30 kBdist/assets/index-Bgopu29T.js 142.18 kB │ gzip: 45.71 kB✓ built in 1.82s▸validating in headless Chromium… 200 OK, 0 errors▸live at shop.autodeploy.fyi
02 · recover
Failures, handled.
Known failures are fixed by deterministic rules in seconds — wrong Node version, wrong output folder, stale lockfile. The rest go to a model that proposes the smallest patch, which is policy-checked, rebuilt in a fresh sandbox and validated before you ever see it. Nothing touches your main branch without you.
03 · release
Every deploy is on the record.
Builds are content-addressed and immutable. Going live — or going back — moves one pointer at the edge. No rebuild, no downtime, no guessing which files are served.
rollback #128 → #127 · 3.7 ms · nothing rebuilt
04 · proof
Numbers we measured.
Not projections. Each figure comes from a test or a run you can reproduce from the repository.
- duplicate job claims
- 0
- 500 jobs raced by 8 workers (integration test)
- failures auto-recovered
- 82%
- 38-repo failure corpus · 64% by rules alone · 0 false fixes
- worker killed → rebuild
- 36s
- SIGKILL mid-build; lease expiry + reaper (chaos test)
- rollback
- 3.7ms
- pointer flip under an advisory lock, nothing rebuilt
05 · system
Under the hood.
Jobs live in Postgres, where leases and retries belong. Events fan out through Kafka to independent consumers. Static files are served from the edge; server-rendered apps wake on demand in sandboxed containers.
- 01GitHubA push or pull request arrives as a webhook.
- 02APIFastify checks it and records the deployment and its job in one transaction.
- 03PostgresHolds the job queue and every state change; events fan out to Kafka.
- 04WorkerClaims the job under a lease and builds in a gVisor sandbox.
- 05AI gatewayWhen rules can’t fix a failure, proposes a patch for a person to approve.
- 06EdgeServes the result from Cloudflare: files from B2, server apps on demand.
Ship something broken.
We'll tell you why — and most of the time, fix it before you notice.