Machine Control

Open source · active pre-1.0

Give your coding agent a cross-platform test lab.

Build on one machine. Let the agent launch, inspect, and verify the real application across Windows, macOS, and Linux VMs through one compact, target-native interface.

Read the quick start View source
  • Windows · macOS · Linux
  • Native UI semantics
  • Local or remote
  • MIT licensed
application check3 targets
$ mc targets
windows   ready   windows.uia
macos     ready   macos.ax
linux     ready   linux.atspi

$ mc --target windows desktop snapshot \
    --query Save --projection compact

route          windows.uia
delivery       confirmed
effect         confirmed
host impact    none

One workflow, native controls

The interface is shared. The platform stays real.

Machine Control normalizes target selection, observations, actions, and results without pretending that every operating system works the same way.

The development loop

From a code change to evidence on every target.

An agent or CI system chooses the work. Machine Control provides the safe, repeatable boundary to each configured test machine.

  1. 01
    Discover

    List eligible targets and inspect their real capabilities.

  2. 02
    Claim

    Reserve an exact VM or acquire an isolated workspace.

  3. 03
    Deploy

    Build, install, launch, and prepare the application.

  4. 04
    Exercise

    Observe compact semantics and act through native controls.

  5. 05
    Verify

    Separate action delivery from independently observed effect.

Context is part of the system

Send useful state, not an endless desktop stream.

Compact semantic projections keep stable roles, labels, values, bounds, and actions while filtering provider noise. Content digests let an unchanged observation return without repeating the element tree.

These are measured response bytes from current Windows fixtures, not a claim about exact model-token savings. A cross-provider token benchmark is planned.

Read the measurements →
Calculatorresponse bytes
full9,971
compact6,577
unchanged1,165
Settingsresponse bytes
full4,971
compact3,843
unchanged986

What it solves

A machine layer for cross-platform application work.

01 / ONE VOCABULARY

Change the target, not the agent workflow.

Applications, windows, semantic snapshots, captures, workspaces, and results share a stable ergonomic direction across platforms.

02 / NATIVE DEPTH

Keep the operating system’s strongest controls.

UIA, Accessibility, AT-SPI, XCTest, ADB, and platform services remain real providers rather than disappearing behind pixels.

03 / QUIET OPERATION

Test the target without taking over the host.

Ordinary operations execute inside the target. Hypervisor windows and host input stay explicit recovery routes.

04 / TRUTHFUL RESULTS

Know whether the application actually changed.

Acceptance, delivery, effect, evidence, provider route, and uncertainty are reported separately.

Decision guideSee where Machine Control fits beside Playwright, Appium, Cua, and BrowserStack.
Compare the options →

Built under real application pressure

Powering cross-platform development at Graehl Arts.

Machine Control is used to develop and exercise applications across desktop and device targets—not only synthetic automation fixtures.

Visit Graehl Arts →
JSTorrentShared TypeScript product across desktop, browser, Android, ChromeOS, and iOS.RSTorrentRust engine with maintained desktop, Android, and iOS clients.AtpianoNative macOS and Windows application packaging and UI workflows.200 OK Web ServerCross-platform web-server application for desktop, Android, and ChromeOS.

Current evidence

Broad destination. Honest status.

The common desktop contract is exercised today. Device families retain explicit native operations while broader facade coverage is still growing.

TargetCurrent routeStatus
WindowsCommon desktop · UIA / Cua / Win32First complete vertical slice
macOSCommon desktop · AX / Quartz / CuaAccepted VM appliance
LinuxCommon desktop · AT-SPI / WaylandAccepted VM appliance
DevicesChromeOS · iOS · Android · QuestWorking native routes; varied coverage

“Accepted” means exercised against the repository’s conformance and real-application workflows. It is not a production support SLA.Full platform status →

Field notes

Why app-development agents need a machine layer.

The first project note explains why screenshots, native automation, VM lifecycle, and effect verification belong in one development loop—but not one lowest-common-denominator implementation.

Read the note →
01

Architecture

8 minute read

August 25, 2026