Skip to content
Autowright

Introduction

Autowright makes Playwright browser automation self-healing. You change one import line in your script; from that moment every run is invisibly observed. When the script fails — a selector broke, a page changed, a login stopped working — Autowright captures the evidence, diagnoses the failure, and hands it to an AI agent that opens a pull request with the fix.

The problem

Browser automation breaks constantly, and never because your code was wrong. Sites redesign, selectors get renamed, flows change. Every break means a human reproducing the failure, digging through logs, finding the new selector, and shipping a one-line fix. Autowright automates exactly that loop.

What you get

  • Zero-rewrite integration. Your script is plain Playwright. Autowright wraps the browser with a transparent proxy — browser.newPage() still returns a real Playwright Page.
  • Evidence, always. Actions, screenshots, DOM snapshots, network requests, and console logs are captured on every run and bundled when something fails.
  • Fixes as pull requests. A Claude-powered agent clones your repo, reasons over the captured evidence, and opens a PR you review before merging.
  • Never blocked, never silent. Errors are re-thrown to your script, never swallowed. Every failure ends in one of two visible outcomes:
    • Fix proposed — a confident fix, delivered as a PR to verify against the live site.
    • Needs attention — no confident fix; you get the diagnosis and the evidence bundle instead.

The three layers

┌─────────────────┐   error +    ┌─────────────────┐   pending    ┌─────────────────┐
│  Browser Proxy  │  evidence    │     Service     │    jobs      │      Fixer      │
│  (your script)  │ ───────────▶ │  (coordinator)  │ ───────────▶ │   (AI agent)    │
│                 │              │                 │              │                 │
│ capture actions │              │ dedup by hash   │              │ clone repo      │
│ classify errors │              │ track failures  │              │ analyze DOM     │
│ bundle to S3    │              │ queue fixes     │              │ open PR + Slack │
└─────────────────┘              └─────────────────┘              └─────────────────┘
  1. Browser (@autowright/browser) lives in your script and captures everything through a JS proxy around Playwright.
  2. Service (@autowright/service) receives lightweight failure metadata, deduplicates repeat incidents, and queues fix jobs.
  3. Fixer (@autowright/fixer) runs the Claude Agent SDK in a sandbox, produces the fix, opens the PR, and notifies Slack — with the real dollar cost of every fix attached.

Read How It Works for the full data flow, or jump straight to the Quick Start.