GW
macOS app · public beta · free

Your Mac does the work. You just aren't there.

AI Box turns the Mac on your desk into an AI workstation — an assistant that browses, books and builds things for you, a writing studio, image generation, animated ads and explainers, and a real terminal — then hands you the controls on your phone. Tell it what to do, by voice or text, and it keeps working after you walk away. It asks before anything is paid, sent or deleted. The machine does the work. The keys never leave it.

Download for macOSRead the source

Universal — Apple Silicon & Intel · Signed & notarized · macOS 10.15+

This is a public beta. It's stable enough that I use it for my own work every day, and free while it's in testing. What I want back is what broke, what confused you, and what you wish it did — open an issue or email me.

Powerful AI is either in someone else's cloud, or chained to your desk

Cloud tools see everything

Your manuscript, your client's codebase, your API keys. You accept someone else's content policy and retention terms to get work done.

Local tools stop at the desk

Running models on your own hardware fixes privacy and creates a new problem: the moment you stand up, the most capable computer you own goes idle.

AI Box does both

The Mac stays the source of truth and the only holder of secrets. Your phone is a window onto it — same app, same documents, over your own network. Nothing is hosted for you, because nothing needs to be.
New in 0.7

Watch it browse, right beside the chat — and keep working

The assistant's browser now lives in a panel next to the conversation, the way it does in Claude's desktop app. You see every page it opens as it opens it, and you can reach in at any moment: click, scroll, type, paste. Meanwhile its Chrome stays out of your way — it never takes your screen, your mouse or your keyboard, so you get on with something else while it works.

A browser in the app

Back, forward, reload, an address bar, and the page laid out at the panel's own size — not a big page shrunk to fit. Drag the divider or widen it; on the phone it fills the screen and you swipe to scroll.

Sign in once, there

Logins, codes and CAPTCHAs are done by you, in the panel, when the assistant hands over — and they are kept. Sign in to a site before a task and every task after can use it, across restarts.

Full autonomy, if you want it

A clearly marked mode for long jobs: no approval prompts and no step limit, so it runs until the work is done or it needs you. Off by default — and set on the Mac, so a phone can't turn it on.

And a new Video tab

Animated ads and explainers, voiced, scored and animated on your Mac — with your own logo, for a few cents each. See how it works below.

New in 0.6

Assistant Agent: say what you want done, and walk away

“Book a table for two at 8 on Friday.” “Add this domain to my Vercel project.” “Find the invoice from March and put it in the tax folder.” The assistant does the job itself, step by step, on your Mac — and the job runs there, so it keeps going when you close the window or lock your phone. Tell it more while it works; answer when it asks.

A real browser

It has its own Chrome, shown live beside the chat, and reads each page as a numbered list of buttons, links and fields — so it clicks what is there, not a guess at a coordinate. Ask, and it works in the Brave, Chrome or Safari tab you already have open.

Your Mac, when you ask

The terminal and your files, system switches like Wi-Fi and volume, and — only when you tell it to — your mouse and keyboard, looking at the screen after every step.

You stay in charge

Anything that pays, books, sends or deletes waits for your OK, with a screenshot — on the Mac or on your phone — unless you switch on full autonomy. It never types a password; logins and CAPTCHAs are handed back to you.

Talk to it

Tap the microphone and speak: what you said is sent straight away and the answer is read back. “Yes” or “no” answers its approval cards, in English or German. Type instead, and it answers in text. On the phone this works over the secure Tailscale link below.

New in 0.4

Screen Assist: hold a key, ask out loud, and it does it

Everything above lives in the app's window. This does not. Hold ⌥ Space anywhere on your Mac, say what you want, and let go. A bar appears over whatever you were using, answers you out loud, circles the thing it is talking about — and, if you asked for something to be done, does it.

It sees

A screenshot goes with the question, so “what does this error mean” needs no explaining. The answer is spoken, and the thing it refers to is circled on your actual screen — over any app, full-screen ones included.

It hears

Hold the key, speak, let go. The audio goes to the model alongside the screenshot in one round trip — it is not transcribed first and then asked, which is the delay you would otherwise feel every single time.

It acts

“Turn Bluetooth off.” “Open Mail.” “Scroll down and click Accept.” It clicks and types for real, one step at a time, looking at a fresh screenshot after each one before deciding the next.

Two kinds of hands, and the difference matters

System switches

Bluetooth, Wi-Fi, volume, mute, light and dark, opening and quitting apps, opening a link, locking the screen. These go straight to macOS — exact, instant, and no window has to be open. It is told to prefer them over clicking.

Real clicks and keystrokes

Everything macOS gives no API for. Events are synthesised at the same level as your own mouse and keyboard, so they work in every app — native, Electron, a web page — with no per-app automation support and no extensions.

You watch it happen, and you can stop it

  • ▪Every step is named as it happens — “Clicking the Wi-Fi menu” — and ticked or flagged as it finishes
  • ▪Escape stops it from anywhere on the Mac, even once the app it is driving has the keyboard
  • ▪A ceiling on actions per request, or no ceiling at all if you are watching and want a long job finished
  • ▪It will not buy, send, post or delete anything, or touch passwords, keychains or payment details
  • ▪The bar floats without bringing AI Box forward — your work stays exactly where it was
  • ▪Answers are spoken in the language you asked in, by a voice that actually speaks it

Only models that can actually do the job

This needs one model that can see a screen, hear a spoken question and call tools. Most cannot. The picker lists only those that can, and a model you choose is sent a real request — a screenshot, a spoken sentence and a tool — before it is kept. Anything that refuses, stalls or answers in the wrong shape is dropped and never offered again, so a failure lands here rather than mid-question.

There is a local option too: a vision model on your own Mac, through Ollama, with the screenshot never leaving the machine. Slower, and typed rather than spoken — no local model has ears yet — but nothing is uploaded at all.

macOS asks your permission for each part the first time it is needed — screen recording to see, the microphone to hear, accessibility to click. Nothing is granted quietly, and any of them can be withdrawn without breaking the rest.

Five tabs down the left. This is what each one does.

They share one window, one model picker and one workspace — so a passage you wrote can be illustrated, and a file the agent touched is the same file your terminal is sitting in.

The Assistant Agent tab in AI Box, with the conversation and composer
01Assistant Agent

An assistant with hands, not a chat box

Ask a question and it answers. Ask it to do something — book, find, fix, build — and it does it on the machine rather than describing what you should do.

  • Browses in a real Chrome, works in your open tabs, runs shell commands, reads and edits files
  • Its browser live in a panel beside the chat — watch it, or click and type in it yourself
  • Steps and approval cards right in the conversation
  • Drop or paste images in; speak or type; carry on the conversation after it finishes
The Write tab in AI Box, showing a manuscript, word count and the outline panel
02Write

A manuscript editor that finishes your sentence

A proper writing surface rather than a chat transcript — your document stays the document, and the AI works inside it. No draft is ever really gone.

  • Continue a passage mid-sentence, or proofread a selection and accept corrections one at a time
  • Modes for fiction, business, marketing and academic work
  • Chapter and scene outline, word goals, and version history on disk
The Images tab in AI Box, where local image generation runs
03Images

Your own GPU, or the best cloud models

Run it locally and nothing is uploaded to be rendered — the models run on the machine under your desk. Or reach for a cloud model when you want one, on your own key.

  • A managed ComfyUI the app installs for you — no Python setup
  • Or a cloud model on your chosen provider's key — Google's Nano Banana on Gemini, GPT Image on OpenAI, or anything on OpenRouter
  • Exact output sizes — ask for 2400 × 2400 and get exactly that, resampled on the Mac
  • Illustrate a scene straight from the passage you have selected in Write
The Video tab in AI Box: the writing and image models, the voice, format, a length slider, the brief and your own images
04Video

Animated ads and explainers, made on your Mac

Motion graphics, not stock footage: cut-out images, kinetic type and your brand, animated in code and filmed frame by frame into an MP4. Nothing paid runs until you press Make video — and then it's a few minutes and a few cents.

  • A voice that drives the picture — any speech model on OpenRouter, or ElevenLabs on your own key — with every word timed and every animation placed on it
  • Your logo and photos used as they are, transparency included; paste an App Store link and the real icon and screenshots are used
  • Pick the models, the voice (hear a sample first), the format and a length from 10 seconds to 2 minutes; “Write script” drafts the voice-over for you to edit
  • Music and effects synthesised on the Mac, mixed under the voice to −14 LUFS — no licences to worry about
  • “Make the logo bigger” redoes only the animation, in under a minute for about a cent — and every version is kept
The Terminal tab in AI Box, a real PTY session
05Terminal

A real terminal, not a command box

The tab that makes the phone useful: a genuine shell on your Mac, in your pocket, with the same session waiting when you sit back down.

  • An actual PTY — vim, top and claude all behave normally
  • xterm.js on the front, Rust PTY sessions behind it
  • Reachable from the sofa, or from a train

Your key, your account, your choice of provider

Most apps pick a provider for you and resell you the tokens. AI Box asks which one you want and then gets out of the way.

OpenRouter

One key, every model — Claude, GPT, Gemini, Grok, Qwen, the lot, read live from the catalogue so a model released this week is there without an update.

Anthropic, OpenAI or Google — directly

Credit already sitting on an account you pay for is reachable on its own key, billed by them with nothing in between. Anthropic goes through its own Messages API rather than a compatibility shim, so tool use behaves properly instead of breaking at the first tool call.

One at a time, and you choose it

Whichever provider you pick is what every tab uses — the assistant, Write, Images and Video. No merged lists, so the same model never appears twice at two different prices, and there is never a question about which account just paid. The other keys stay saved — switching back is one click.

Or no provider at all

Models installed through Ollama are offered whichever provider is active, because nothing about them is billed. That is also the fallback when the internet is gone or the credit has run out.

Three steps, and one of them is optional

  • 01

    Install it

    Download the .dmg and drag it across. It's signed and notarized by Apple, so it opens like any other app — no right-click-to-open dance. It updates itself from then on.

  • 02

    Pick a provider, paste one key

    Choose OpenRouter for one key that reaches every model, or go straight to Anthropic, OpenAI or Google and spend credit you already have there. Whichever you pick is what every model picker in the app offers — no mixed lists, no wondering which account paid. Or install a model through Ollama from inside the app and run with no key at all. Keys go into the macOS Keychain, never a config file.

  • 03

    Pair your phone — optional

    Scan a QR code. Your phone opens the same app — over your home network, or anywhere through Tailscale — and installs to the home screen.

At home it just works. Away from home takes one more app.

AI Box never puts your Mac on the public internet, so how you reach it depends on where you are. Both routes carry the same pairing token, and neither opens a port to the world.

On your own network — nothing to install

The Mac serves the app to your Wi‑Fi. Scan the QR code in Settings → Phone access and your phone is paired. That's the whole setup, and it covers the sofa, the kitchen table and the garden.

Anywhere else — add Tailscale

Install Tailscale on the Mac and on your phone, sign both into the same account, and they join one private network of your own. Turn on HTTPS in Settings → Devices and scan that QR code: the Mac is reachable from a train, an office, another country, over a secure link with a real certificate — which is also what lets the phone use its microphone. Nothing is exposed publicly. Free for personal use.

Leave the Mac awake

A sleeping Mac answers nothing. Turn on Away mode before you leave and AI Box holds the machine awake — display off, work continuing — for as long as it's serving.

One codebase, two transports

Every capability is written once against a transport layer that resolves at runtime to either Tauri IPC or an HTTP + WebSocket connection. That is why the phone isn't a cut-down companion app — it is the same app.

ShellTauri 2 — a Rust backend with the system WebView. ~17 MB installed, not a bundled browser.
InterfaceReact 19 + TypeScript. TipTap/ProseMirror for the editor, xterm.js for the terminal.
BackendRust: provider streaming with server-side cancellation, a task runner that keeps the assistant going with no window open, Chrome over the DevTools protocol, PTY sessions, filesystem tools behind a path guard.
RemoteAn axum HTTP + WebSocket server on the Mac serving the same bundle to the phone. Bearer auth, constant-time comparison, fails closed.
SyncItem-level merge with tombstones. Edit the same document on both devices and neither side loses work.
DistributionUniversal binary, Developer ID signed, Apple-notarized, with minisign-verified auto-updates.

An agent with real access needs real limits

  • Your API key never leaves the Mac. The phone gets a placeholder so its interface knows a key exists; the Mac injects the real one when the request goes out.
  • File access is jailed to a workspace folder you choose, with .. traversal and symlink escapes rejected in Rust, not in the interface.
  • Credentials are off-limits, always. ~/.ssh, ~/.aws, keychains, shell startup files and LaunchAgents are refused outright — and turning off approval prompts does not turn this off.
  • Nothing final happens without you. Paying, booking, sending and deleting wait for your approval, on whichever device you are using. What counts as final is read from the page itself, not decided by the model — a web page can talk a model into things, but it cannot rename the button. Full autonomy lifts this, and only when you switch it on, on the Mac.
  • Passwords stay yours. The assistant is never shown a password field, so it cannot type into one; logins, codes and CAPTCHAs are always handed back to you.
  • Your words stay yours. Drafts are snapshotted to disk continuously, restorable, and exportable to a folder you own. With a local model, nothing leaves the machine at all.
Get the free beta

Take it for a run, then tell me what broke

AI Box is free during beta and never hosts your content or pays for inference — you choose the model and the account it bills to.

Download for macOSSend feedback

Designed and built by GW‑InTech — available for work.