macOS + Linux
Two engines. One contract.
Headless uses WKWebView on macOS and sandboxed Chromium on Linux. Capability differences are explicit and unsupported operations return errors instead of partial behavior.
01 / Supported hosts
Native where it runs.
macOS
macOS 13+: visible WKWebView windows.
Linux
Linux (Ubuntu and other distros): sandboxed Chromium, headless by default or visible with DISPLAY.
02 / Computer-use comparison
Choose for the actual job.
Scores are qualitative capability judgments from the README, not benchmark measurements.
| Dimension | Coordinate CU (regular browser) | Scripted (PW / Puppeteer / Selenium) | Headless |
|---|---|---|---|
| Targeting precision | 2 | 4 | 5 |
| Safety / blast radius | 2 | 3 | 5 |
| Evidence (shots, video, QA) | 3 | 3 | 5 |
| Agent surface (tokens / glue) | 2 | 3 | 5 |
| Setup friction | 3 | 3 | 4 |
| Platform coverage (host OS) | 5 | 5 | 4 |
| Desktop / OS reach | 5 | 1 | 1 |
- Coordinate CU — strong when the agent needs the whole desktop; weaker on precise web targeting (pixels drift), larger screenshot/prompt cost, and a broader OS blast radius. Runs on Windows, macOS, and Linux.
- Scripted — reliable selectors and driver APIs; more agent glue to write and maintain; evidence and security posture are usually bolted on. Broad host-OS support including Windows.
- Headless — semantic role/name/@ref actions, private socket + allowlisted commands, built-in record/screenshot/diagnostics; browser-only (no desktop). Runs on macOS and Linux (Ubuntu and common CI/test distros); Windows is the exception.