Your work begins and ends inside a web page.
Use Playwright. Pair it with Machine Control when the same test must also install an app, handle native dialogs, exercise an updater, or verify behavior on several operating systems.
Comparison guide · reviewed August 25, 2026
Different layers. Different jobs.
Machine Control is the target layer between an agent or CI job and the strongest controls available on each machine. It does not replace a browser test framework, a mobile driver, a computer-use provider, or a hosted device cloud. It makes those pieces usable as one governed cross-platform test lab.
Notan autonomous agent
Notone universal UI driver
It isthe machine-facing contract
Where it sits
Keep specialized depth. Standardize the development loop.
The caller selects and claims a target once. Machine Control then routes each operation to a native or upstream provider and returns the actual route and observed outcome.
CALLERAgent or CI
TARGET CONTRACTMachine Controlinventory · doctor · claim · route · verify
WebPlaywright · CDP
DesktopCua · UIA · AX · AT-SPI
DevicesXCTest · ADB · Appium
Windows and macOS VMs · Linux desktops · ChromeOS · iOS · Android · Quest
Capability matrix
What each option owns.
This compares documented first-class scope, not every feature that could be added with custom code. A partial mark means the capability is narrower, provider-dependent, or belongs to an adjacent product.
✓ First-class◐ Partial or narrower↔ Pair another layer— Not its scope
| Capability | Machine Controltarget layer | Playwrightweb testing | Appiumdriver framework | Cuacomputer use | BrowserStacktest cloud |
|---|---|---|---|---|---|
| Web-page DOM and browser assertions | ↔Pair itUse Playwright or CDP for page-level depth. | ✓CoreDOM locators, assertions, traces, and browser projects. | ◐Driver-basedBrowsers and WebViews through installed drivers. | ◐AvailableBrowser semantics, pixels, and a page escalation route. | ✓HostedRuns Playwright, Selenium, Cypress, and related frameworks. |
| Native mobile application automation | ◐GrowingWorking iOS and Android routes; common coverage varies. | —Mobile webDevice profiles emulate browser viewports, not native apps. | ✓CoreXCUITest, UiAutomator2, Espresso, and other drivers. | ◐AvailableAndroid sandboxes and optional device tooling. | ✓CoreReal iOS and Android devices with native test frameworks. |
| Desktop applications and operating-system UI | ✓CoreWindows, macOS, Linux, and system surfaces. | ◐ElectronExperimental Electron API; otherwise browser-focused. | ◐Driver-basedOfficial Mac2 and Windows drivers; other coverage varies. | ✓CoreReal Windows, macOS, and Linux desktops. | —Not its focusIts documented products focus on web and mobile testing. |
| Hosted, disposable machines or devices | —Bring your ownMachine Control is not a hosted test-lab service. | —Bring a runnerA framework, not a machine fleet. | —Bring infrastructureA server and driver ecosystem, not a device cloud. | ✓CoreCloud and local computer sandboxes. | ✓CoreManaged browser and real-device capacity. |
| Target inventory, readiness, and lifecycle | ✓CoreLogical targets, read-only doctor, start, stop, and workspaces. | ◐Browser scopeLaunches or attaches to browsers, not whole targets. | ◐Session scopeCapabilities, app install, and device actions vary by driver. | ✓CoreDriver doctor plus image and sandbox lifecycle. | ✓ManagedA live catalog, sessions, queues, and plan capacity. |
| Exclusive use of a shared existing target | ✓CoreExpiring claims bind exact targets and workspaces. | —Not its scopeParallelism is organized around browser workers. | —Not its scopeInfrastructure decides how a device is allocated. | ◐Fleet claimsNamed claims allocate Cua-managed sandbox pools. | ◐Session slotsManaged allocation and queues, not a bring-your-own lease contract. |
| Semantic UI state plus pixels | ✓UnifiedOS semantics, screenshots, windows, and stable target vocabulary. | ◐Web scopeDOM and ARIA locators plus page screenshots. | ✓Driver-basedElement models and screenshots through the selected driver. | ✓UnifiedAccessibility trees and screenshots in one desktop observation. | ◐Framework-basedUnderlying tools plus visual logs and session media. |
| Compact repeated observations for agent context | ✓Built inCompact projections and unchanged-content digests. | —Different modelLocators query a page rather than returning machine state. | —No common digestPage source and elements remain driver-specific. | ◐SnapshotsLightweight and full snapshots; no shared unchanged digest. | —Different modelLogs, videos, and results are retained as test evidence. |
| Route, delivery, effect, and evidence in action results | ✓Across providersAlso reports attempts, uncertainty, fallback, and host impact. | ◐In test codeActions, assertions, and traces are separate primitives. | ◐In test codeDriver response and postcondition assertions stay separate. | ✓Driver contractEffect, route, delivery, evidence, and escalation are explicit. | ◐ObservabilitySession logs, screenshots, video, and test status. |
| Provider arbitration and explicit safe escalation | ✓CoreRoutes each operation across Cua and native providers. | —Single layerBrowser engines are test targets, not fallback providers. | ◐Choose a driverThe caller selects the driver before the session. | ✓Action ladderAccessibility, pixels, foreground, and page escalation. | ◐Choose a frameworkRuns several frameworks without unifying their contracts. |
| Open-source, self-hostable core | ✓MITPublic contracts, implementation, fixtures, and evidence. | ✓Apache-2.0Open-source browser automation and test runner. | ✓Apache-2.0Open-source server, clients, drivers, and plugins. | ✓MITOpen-source core with optional hosted services. | —CommercialA self-hosted browser-grid product is available commercially. |
“Supported” is not a universal quality score. Playwright is much deeper inside a web page; Appium is much more established for native mobile test suites; Cua already supplies several excellent desktop and sandbox primitives; BrowserStack supplies infrastructure Machine Control does not attempt to sell.
The practical choice
Choose by the boundary your test crosses.
You are building a conventional native mobile test suite.
Use Appium and its platform drivers. Machine Control’s device routes are for bringing lifecycle, target selection, evidence, and agent ergonomics into the larger cross-platform loop.
You need excellent computer-use primitives or fresh sandboxes.
Use Cua directly when that is the whole job. Machine Control already uses pinned Cua Driver routes where they win, then adds native adapters and a broader target, lifecycle, policy, and conformance boundary.
You want to buy broad browser and real-device capacity.
Use BrowserStack. Machine Control is aimed at an open, agent-facing contract over machines and devices you operate or explicitly connect—not a replacement commercial device cloud.
The difference in one sentence
Machine Control governs the whole target, not only the next click.
Resolve the exact target, inspect readiness without mutation, acquire exclusive use, and start only through an authorized lifecycle route.
Use native semantics and pixels through the best measured provider without silently switching to a disruptive host-side route.
Report acceptance, delivery, observed effect, evidence, uncertainty, and cleanup ownership as different facts.
Honest limits
Broad destination. Uneven maturity.
- Machine Control is active and pre-1.0; schemas and setup are still evolving.
- Windows is the first complete vertical slice. macOS and Linux have accepted VM appliances.
- ChromeOS, iOS, Android, Quest, and Steam Deck have working native routes with varied common-facade coverage.
- You currently supply and configure the machines, hypervisors, and physical devices.
- Use Playwright for deep browser tests and Appium or native frameworks for mature mobile suites.
Sources and methodology
Check the claims.
The matrix was reviewed against each project’s official documentation and Machine Control’s checked-in implementation evidence. Products evolve; corrections are welcome.
Machine Control
Current platform statusProvider landscapeCapabilities and resultsPlaywright
Browsers and projectsLocatorsElectron APIAppium
Appium in a nutshellHow drivers workDriver ecosystemCua
Sandboxes and real machinesCua Driver contractsSandbox pool claimsBrowserStack
Product documentationBrowser and device selectionApp Automate evidenceProduct names and trademarks belong to their respective owners. This project is not affiliated with or endorsed by them.