GHOST HANDS / COMPARED
Ghost Hands, honestly compared
Last updated
Four tools get picked when software has to drive a browser: Ghost Hands, Playwright, Stagehand, and Browser Use. Three of them are bigger, older, and better backed than the fourth, and on several rows below they simply win. Every claim about them on this page was checked against their own documentation on November 11, 2026, and where a tool wins, the page says so.
The table
| Ghost Hands | Playwright | Stagehand | Browser Use | |
|---|---|---|---|---|
| What it is for, and who decides each step | Governed agent hands. A decider proposes each action, but the Governor classifies it (readonly, write, or consequential), and policy, a domain allowlist, and human approval decide whether it runs. | A test and automation framework. Your code decides every step; there is no agent and no model in the loop unless you add one yourself. | An AI layer on top of Playwright. The model decides at the steps where you call its AI actions, and your code decides everywhere else, so the two mix line by line. | An autonomous agent. You hand it a task in plain language and the model plans and drives the whole run. |
| How elements are targeted | Numbered element maps from a walk of the page’s composed DOM tree, including shadow roots and iframes. No screenshots and no vision model. | Locators and selectors your code names in advance, with auto-waiting and strict matching. | Plain-language instructions resolved by the model at runtime, such as “click the checkout button”. | An indexed element list from the page that the model picks from at each step. |
| Determinism | A live run uses a decider, but a finished run graduates into a deterministic, model-free script, and since 1.0.0 a proven run also replays from a governed plan cache with no model calls. | Fully deterministic. The same script does the same thing on every run. | Deterministic where you write plain Playwright and model-dependent where you call the AI actions. Its action cache can replay a step without the model, but nothing classifies or approves that replay. | The model stays in the loop for the whole task, so runs vary with the model and with the page. |
| Approval before a consequential action | Built in. Submits, payments, sends, deletes, publishes, and transfers are classified consequential and wait for a human approval, and silence counts as a no. | No approval gate exists in the framework. Anything your script does, it does. | No built-in gate. Its own cookbook shows you writing the pause and the confirmation in your application code. | The library describes its own safety guards as “not a sandbox”. The consequential clicks are the agent’s to make. |
| The record of a run | A JSONL trail written before each action executes, recording the verdict as well as the act. Since 1.0.0 the trail is hash-chained, and the verify-trail command proves it has not been edited. |
Traces, video, and screenshots captured during the run and reviewed afterwards in the Trace Viewer. | Logs and run history from the SDK and from the Browserbase cloud. | A step-by-step history of the agent’s actions and reasoning. |
| Browsers, languages, license, and the upsell | Chromium only, through its own zero-dependency CDP client. Python. MIT license, free, no hosted tier. First published in October 2026 by a one-person studio, and its GitHub star count is zero. | Chromium, Firefox, and WebKit. JavaScript, TypeScript, Python, Java, and .NET. Apache 2.0, free, backed by Microsoft, with roughly 100,000 GitHub stars. | Chromium through Playwright. TypeScript, Python, and Go. MIT, free, with the Browserbase cloud as the paid hosted tier. | Chromium, with stealth and proxy options built for difficult sites. Python. MIT, free as a library, with a hosted cloud and agent API on top; its README prices browser time at $0.02 an hour. Roughly 121,000 GitHub stars. |
Where they win
- Playwright wins on reach and tooling: three real browser engines, a complete test framework around the automation, and the Trace Viewer for debugging. If you are testing your own site, use Playwright.
- Stagehand wins on flexibility: three languages, AI exactly at the steps you choose instead of everywhere or nowhere, and a managed cloud of browsers so you never run Chromium yourself.
- Browser Use wins on autonomy: it is built to push through the modern web’s defenses with stealth browsers, proxies, and CAPTCHA handling, and a community of roughly 121,000 stars stands behind it.
- Ghost Hands wins the governance rows: classification and approval before anything consequential, a record written before the act and tamper-evident since 1.0.0, and finished runs that stop needing a model. Those are the rows it was built for, and they are the rows the others leave open.
Honest limits
There is no speed row in the table. Ghost Hands has never been benchmarked against the others, so there is no honest number to print. There is no success-rate row either: no fair head-to-head test exists, and inventing one would be marketing, not comparison.
Ghost Hands, stated as plainly as its own page states it: a real model has driven the graded local fixture suite (3 of 4 tasks passed on November 12, 2026, the failure published), while driving live public sites end to end remains unproven; the Chromium driver has been proven live on two sites, example.com and Wikipedia; the phone body is pending on-device proof; and its hosted bus mode is proven only against a locally booted server.
Sources: checked November 11, 2026, against playwright.dev, the Stagehand documentation, the Browser Use documentation and GitHub repository, and the Ghost Hands repository. Star counts are the projects’ own GitHub figures on that date and move over time.
Questions? See the Hub help center.