Windows · macOS · Android · iOS · Browser · Linux

Give AI agents
real hands.

Handox is the cross-platform effector layer for AI agents — the ability to click, type, swipe, and read the screen across every device, through one binary and one API.

the motor cortex for the agentic era — LLMs think, Handox does

Windows
macOS
Android
iOS
Browser
Linux
The problem

Agents can think. They can't do.

LLMs reason, plan, and decide — but when it's time to actually operate a computer or a phone, they hit a wall. Device control is fragmented: ADB for Android, AppleScript for Mac, Win32 for Windows, Playwright for browsers. None of them talk to each other, and every agent developer reinvents it, poorly, one platform at a time.

Today, without Handox

Five different tools, five mental models, five brittle integrations — each cloud-dependent, model-specific, or locked to one environment.

adb shell input tap …
osascript -e 'tell app …'
SendInput(…) // Win32
await page.click(…) // Playwright
↳ none of them compose

With Handox

One installable binary. One opinionated API. Every platform behaves identically — if a capability is on the trait, every driver supports it or says exactly why not.

handox.screenshot(device)
handox.click(device, x, y)
handox.type_text(device, "…")
handox.launch_app(device, "…")
↳ same call, any device
Why Handox

Built for agents that act in the real world.

One API, every device

A single DeviceDriver trait — Windows, macOS, Android, iOS, browser, Linux. Cross-platform parity is the whole point.

Local-first

Runs on your machine, not someone's cloud. No model lock-in, no per-action fees, no data leaving the device.

Permissioned & audited

A default-deny permission engine gates every action; an append-only audit log records every click, keystroke, and tool call.

Remote reach

Drive a device on another machine across NAT and the internet through an end-to-end-encrypted relay. No inbound ports.

MCP-native

Speaks the Model Context Protocol out of the box — drop it into Claude, Cowork, or Claude Code and the tools just appear.

Sees & remembers

Native OCR (Vision / WinRT / Tesseract), fuzzy element matching, screenshot annotation, visual diff, and action record/replay.

How it works

Three ways in. One surface out.

Point your agent at Handox over MCP, REST, or the CLI. Every call flows through one router — permission-checked, audited, dispatched to the right driver.

01 — INSTALL

One binary

Build the handox binary and drop it at a stable path. No agents, no daemons to babysit.

02 — CONNECT

Wire it in

Add it as an MCP server, start the REST API + dashboard, or call it one-shot from the CLI.

03 — DRIVE

Take action

Screenshot, click, type, swipe, launch apps, read the UI tree — on the host, a phone, a browser, or a device across the world.

claude_desktop_config.json
// Wire Handox into any MCP client — the tools appear instantly.
{
  "mcpServers": {
    "handox": { "command": "handox", "args": ["mcp"] }
  }
}

# …or start the REST API + fleet dashboard:
$ handox start          # → http://127.0.0.1:7600/
$ handox screenshot -o shot.png   # one-shot from the CLI
The system palette

Five kinds of effector. One language.

Handox organises everything an agent can touch into five effector categories — the colour system that runs through the CLI, the dashboard, and the mark itself.

files

Local filesystem

web

Browser & HTTP

shell

OS & process

apis

Third-party APIs

data

Database & storage

Ready to give your agent hands?

Open-source under Apache-2.0. Clone it, build it, and start driving.

Get Handox on GitHub → Read the docs