00demo

AI that sees,
speaks and acts.

Hacking Monkey is a desktop agent for macOS. Talk to it in real time; it looks at your screen, works your apps, and runs multi-step jobs through Claude Code and Codex — on your own API keys.

standby

Wake
Hear
Decide
Act
Verify

Download for macOS Source on GitHub

Apple silicon · macOS 13+ · bring your own API keys

01chain

One spoken sentence, five stages

Every turn runs the same loop. Nothing is faked between stages — if a step can't be verified, it says so instead of claiming success.

Wake — an offline Vosk model on your Mac listens for the wake word. No audio leaves the machine until you're talking to it.
Hear — realtime speech, full duplex. Interrupt it mid-sentence and it stops, like a person would.
Decide — it reads intent from your words, not keywords. Clear instruction: it just does it. Ambiguous or irreversible: it asks first.
Act — commands first, GUI only when there's no command path. Real cursor and keyboard events, 1:1 logical coordinates.
Verify — it looks again and checks the result before telling you it's done.

02does

What it does

Talks in real time

Full-duplex speech. Interrupt it mid-sentence and it stops, like a person would. Voice stays live while work is running, so you can ask what's happening or change your mind without waiting.

Runs your Mac

Opens apps, fills forms, handles dialogs, navigates the browser, changes settings. Real cursor and keyboard events at 1:1 logical coordinates — it works the interface the way you would.

Drives coding agents

Hands engineering work to Claude Code or Codex, tracks them, and reports back out loud. Or connect straight through and talk to an agent directly.

Finishes multi-step work

Give it a goal, not a click. It plans, acts, checks the result, and changes approach when a step fails — instead of stopping at the first obstacle or claiming it's done.

Uses your own keys

Your OpenAI and Anthropic keys stay on your Mac and requests go straight to the providers. You pay them directly; there is no subscription and no server of ours in between.

03data

Measured, not demoed

Demo videos are easy to cut. So we built a 29-task benchmark — 20 command and file tasks, 9 long-chain GUI tasks — with programmatic checkers. The score is what the checker found on disk and on screen, never what the model said it did.

Same machine, same loop, same tasks, same timeout — only the model changes. A "false success" is a task the model reported complete that the checker found incomplete. Full method and per-task logs are in the repo.
Model	Basic 20	GUI chain 9	Total	Wall clock	False success
Claude Sonnet 5	20/20	9/9	29/29	9 min	0
Qwen3.7-Plus	20/20	5/9	25/29	22 min	2
Kimi K3	18/20	4/9	22/29	95 min	0

Ships on Claude Sonnet 5. The others are swappable with one config line — the loop doesn't care which brain it's running.

04data-out

What leaves your machine

Stated plainly, because a thing that watches your screen owes you this.

Stays local — wake-word detection. The offline model runs on your Mac; microphone audio is not sent anywhere until you've woken it and are speaking to it.
Goes out while in use — your speech, and screenshots of your screen, go to the model providers you configured (OpenAI for voice, Anthropic for computer control) to be understood. That is how it sees.
Never collected by us — there is no Hacking Monkey server. Your keys stay on your Mac, requests go straight from your machine to the providers, and we receive nothing.

The camera is different: it only ever opens when you ask for it out loud, enforced in code against the actual transcript — not left to the model's judgment.

05install

Get it

Signed and notarized for macOS. Bring your own OpenAI and Anthropic keys — you pay the providers directly, there's nothing in between.

Download .dmg GitHub Hugging Face

Apple silicon · macOS 13 Ventura or later · needs Screen Recording and Accessibility permission to drive your apps.

AI that sees,speaks and acts.