Most Mac voice control AI tools talk back to you. Atticus does things for you. Hold Control-Option-Space, say what you want, let go — and it opens apps, moves windows, searches files, reads what’s on your screen, fills in forms, even drafts replies to the email you have open. It launched on Product Hunt on 14 October 2026, it is free and open source (MIT), and it was built by Saarth — a high schooler. Our verdict up front: if you already pay for a Claude Pro or Max plan, Atticus is the most cost-effective voice layer for a Mac we have seen this year. But it asks for a level of trust most assistants don’t, so read the watch-outs before you install it. This review breaks down what it does, how it works, what “free” actually costs, where it beats the built-in alternatives, and who should skip it.
What Atticus does
The interaction model is deliberately simple: one hotkey, one sentence, one result. Say “Open Spotify and play lo-fi.” Say “Find me a desk lamp on Amazon and add it to my cart.” Say “Fill out this form with my info.” Because Atticus can read your screen, it also answers questions about the PDF you have open, or writes a reply to the email you are looking at — without you copy-pasting anything. Crucially, it always asks before it buys, sends, or deletes anything. It types the message; you allow the send.
Two features stand out beyond the headline act. First, learned shortcuts: ask for the same action twice and Atticus remembers it, so the next invocation runs without touching the AI at all — the maker’s example is “Open Calculator and Chess” dropping from 5.5 seconds to 0.4 seconds. Second, a Meeting Copilot that listens on calls and surfaces answers to questions people ask you — a feature that reads as genuinely useful rather than demo candy.
How Mac voice control AI actually works in Atticus
The clever engineering is the two-speed architecture. Simple asks — opening an app, moving a window, playing music, web searches — never touch an AI model at all; the maker claims an app opens in 0.08 seconds. Only the genuinely ambiguous requests go to the language model. This is the right design: it keeps latency near zero for the things you do fifty times a day and reserves the expensive brain for the 5% of requests that need reasoning.
The model bill is the second clever bit. Atticus does not ask for an API key. It runs on the Claude plan you already pay for, through Claude Code — so most asks cost you nothing extra beyond your existing subscription. If you don’t have a Claude plan, you can point it at a Devin plan through the Devin CLI, pick a provider you already configured in Pi, or connect Claude Code to a local model server. The project pegs a typical ask at about 0.18 cents of plan usage — roughly 550 asks per dollar at API list rates. Transcription itself is Apple’s on-device speech recognition, so your voice audio never leaves the Mac.
Pricing: what “free” actually means
The app is genuinely free — MIT-licensed, no tier walls, no per-seat upsell. But “free” has an asterisk worth understanding: the compute rides on a subscription you pay someone else for. The practical cost picture looks like this:
- The app itself: $0, forever, open source. You can read the Swift and Python code on GitHub.
- If you have Claude Pro/Max: asks draw from your existing plan at roughly 0.18 cents per ask — effectively free for normal use.
- If you don’t: you need either a Devin plan, a configured provider in Pi, or a local model server. At API list rates you are looking at around $1 per 550 asks — still trivial.
- Transcription: free and on-device via Apple.
Honest maths: this is the cheapest voice-control setup we have evaluated, because every alternative either charges a subscription or burns your own API key. The hidden cost is not money — it is the access you grant, which we cover below.
Who Atticus is for
Atticus fits three profiles well. Developers and power users on Mac who live in hotkeys and already pay for Claude — for them, it is pure upside. People drowning in admin: form filling, window arrangement across displays, file hunting, and “find that thing I saw” requests are exactly what screen-reading voice control does best. Teams experimenting with agent interfaces: the “hold a key and talk” pattern, plus the confirm-before-action design, is a good reference implementation of how voice UIs should behave when they have real power.
It is less obviously for casual Mac users who rarely touch the Terminal or multitask across windows — the built-in tools (see alternatives below) may cover them fine — and for anyone whose machine handles genuinely sensitive data, where the trust calculus needs a careful look.
Watch-outs and honest limitations
No tool that can see your screen, click your windows, and type for you deserves a free pass. Here is the honest version.
- Deep access is the whole point. Atticus drives your Mac through native accessibility controls — clicking, scrolling, typing inside apps. That is what makes it useful, and it is also the largest trust surface. The project ships a “safe mode” that runs the AI’s commands inside Claude Code’s sandbox (writes restricted to a scratch folder), always blocks dangerous commands like rm, sudo, disk tools and shutdown, and asks before anything deletes, sends, buys, or changes a file. Every exact command lands in an ask log you can audit. Good design — but review it yourself, because it is your machine.
- Young project. The Mac repo was created on 10 October 2026 and has around 214 stars with 5 releases. It is moving fast, which is exciting and also means rough edges. A community Windows port exists (FindlayHi’s atticus-windows, MIT), but it is unofficial.
- Mac-only by design. The core app is Swift. No native Linux or Windows version from the author.
- Screen reading raises questions. Audio never leaves your Mac (Apple’s on-device transcription), but the screen context the AI acts on — what you have open — is part of how it works. Fine for most, worth knowing for regulated work.
- Needs a plan to think. Without Claude/Devin/Pi or a local model, you lose the AI half. Simple instant asks still work, but the magic is pay-walled by someone else’s subscription.
How it compares to the alternatives
The comparison that matters is not other AI chatbots — it is the tools already on your Mac. Siri is free and on-device-ish, but it talks back more than it acts; it cannot read your screen or fill a form. macOS Voice Control and Dictation are surprisingly capable for accessibility-grade control and transcription, run fully on-device, and cost nothing — but they have no language-model reasoning, so “answer this email from the PDF I’m reading” is not in their vocabulary. Voice assistants from big tech (Alexa, Google) control smart homes, not desktops.
Atticus sits in the gap: desktop control like Voice Control, but with an LLM’s understanding, and a cost model that piggybacks on a plan you already buy. It is listed on alternative software directories already, and the honest alternative for the price-conscious is simply Voice Control + Dictation if you never need reasoning. If your main need is dictation inside documents, Atticus’s “dictate anywhere” mode competes with paid dictation apps — at $0.
From our experience: the pattern worth stealing
We build AI automation for clients, so what interests us about Atticus is not just the product — it is the design pattern, and it is one we recommend in workflow automation work constantly: route the trivial away from the model. The cheapest AI call is the one you never make. Atticus runs instant deterministic actions for common asks, caches repeats as learned shortcuts, and only spends model tokens on genuine ambiguity. That is exactly how you keep agent systems fast and cheap in production — most teams do the opposite, routing everything through the LLM and wondering why costs and latency spiral.
The second pattern is confirmation as a first-class design. Voice interfaces that can act are dangerous when they guess silently. Atticus asking before every destructive action, logging every command, and sandboxing file writes is the minimum bar for any agentic tool we would put near a business machine. If you are evaluating any agent tool — for yourself or a team — ask for these three things first: a sandbox, an approval step on irreversible actions, and a full command log.
Who should not install it: shared or family Macs where one person’s “remember…” facts could leak into another’s session, and regulated environments (legal, medical, finance) until someone with a security brief has reviewed the accessibility-permission trust surface. For everyone else on a Mac with a Claude plan, it is a confident try-it-today.
Frequently asked questions
Is Atticus really free?
Yes — the app is MIT-licensed and costs nothing. The AI compute rides on a subscription you already pay for (Claude Pro/Max via Claude Code, a Devin plan, or your own provider), at roughly 0.18 cents of plan usage per ask. Transcription is Apple’s free on-device speech recognition.
Does my voice leave my Mac?
No. Atticus uses Apple’s on-device speech recognition, so your voice audio never leaves the machine. The screen context the AI reads (what you have open) is how it understands your requests — that is the part to be comfortable with.
What permissions does Atticus need?
macOS accessibility permissions, so it can click, scroll, and type inside apps. A safe mode restricts the AI’s commands to Claude Code’s sandbox with a scratch folder for writes, blocks dangerous commands, and requires your approval before anything deletes, sends, buys, or changes a file.
Does Atticus work on Windows or Linux?
Not officially — the author’s app is Swift for macOS. A community Windows port by FindlayHi exists on GitHub (also MIT), with its own integrations such as Home Assistant and Jellyfin, but it is unofficial and we have not evaluated it.
How is this different from just using Siri?
Siri mostly answers; Atticus acts. It reads your screen, moves windows, fills forms, drafts replies to open emails, learns your repeated actions as shortcuts, and always confirms before anything irreversible. If you only need transcription, macOS Dictation remains the free on-device answer.
If voice-first interfaces and agent tooling are on your roadmap, browse our AI Tools Directory for more hands-on reviews — and if you want this kind of automation designed for your own workflows, talk to us or explore our AI integration and automation services.