Canary

Canary is an open-source QA harness purpose-built for coding agents like Claude Code, Cursor, and Codex. It reads code diffs, identifies affected UI flows, and drives real browser instances using a QuickJS WASM sandbox that exposes the full Playwright API. The project is MIT-licensed and hosted on GitHub under the wizenheimer/canary repository.

What It Is

Canary sits at the intersection of browser automation and AI coding agents. Rather than forcing a choice between opaque agent runs you can't reproduce and raw Playwright scripts you have to write and maintain by hand, Canary does both: the agent performs the QA and hands back a reproducible Playwright script. Every session captures screen recordings, console logs, network requests, HAR files, and Playwright traces in a self-contained report.html that requires no server or build step to open.

How the Agent Workflow Works

The core loop is straightforward: describe a UI flow in plain language, and the agent drives a real browser, then returns both a human-readable report and the exact Playwright script behind it. In Claude Code, Canary registers as a first-class plugin with slash commands (/canary:verify, /canary:session, /canary:run, /canary:review) and subagents. Cursor and Codex integrations are also available via their respective plugin marketplaces, all pointing at the same shared skills/, agents/, and commands/ directories.

Architecture: Three Tools, One Runtime

Canary ships as a pnpm + Turborepo monorepo with three user-facing tools sharing a single background daemon:

• @usecanary/cli (canary) — the main session orchestrator; records capture-enabled QA sessions and renders reports
• @usecanary/browser (canary-browser) — a lightweight engine for quick, one-off browser automation with no recording overhead
• @usecanary/ui (canary-viewer) — a local Astro-based viewer for browsing, searching, and replaying recorded sessions

The daemon runs Playwright and a QuickJS WASM sandbox. Scripts execute in a sandboxed environment with no arbitrary host access — no Node.js module system, no direct filesystem or network access from script context, and enforced memory and CPU limits.

What Gets Captured

Every Canary session records a comprehensive evidence trail by default:

• Video replay with a per-step filmstrip and scrubbing
• Playwright trace (trace.zip) decodable with npx playwright show-trace
• Network HAR with per-request headers, payloads, and response inspection
• Console log filterable by level (errors, warnings, info, logs) with source URLs
• Reproducible Playwright script — the exact calls (goto, waitForSelector, evaluate, screenshot) with params and timing
• Self-contained report.html — one file, no server, committable and shareable

Individual capture streams can be disabled with --no-trace, --no-video, --no-har, or --no-console.

Setup Path

Installation requires Node 20+ and pnpm. The quickest path is:

npm i -g @usecanary/cli @usecanary/ui
canary install   # one-time: downloads Chromium + runtime (~150 MB) into ~/.canary

A guided wizard (npm create canary@latest) handles the full setup interactively. All commands also run one-off via npx without a global install. Agent plugin installation uses each agent's own marketplace mechanism — Claude Code via /plugin marketplace add wizenheimer/canary, Cursor via its Marketplace UI, and Codex via codex marketplace add wizenheimer/canary.

Current Status

The repository was created in June 2026 and had 344 stars and 19 forks as of mid-June 2026, with active development reflected in recent pushes. The project is MIT-licensed, with portions derived from MIT-licensed work by Sawyer Hood. The license file in the repository is listed as "NOASSERTION" in GitHub metadata, though the LICENSE file itself is a standard MIT license text.