· via Hacker News – Front Page (native)
Atticus is a free open-source macOS voice assistant that skips AI for simple commands
An open-source hold-to-talk assistant for macOS pairs Apple's on-device speech with Claude Code and routes simple actions around the AI entirely.
An open-source voice assistant named Atticus reached Hacker News's front page on October 12, 2026. Hold Control-Option-Space, say what you need, let go, and the menu-bar app arranges windows, opens apps, searches files, reads your screen and answers questions, with Anthropic's Claude doing the reasoning while simple requests are handled without any AI at all.
How it works
According to the project's GitHub page, Atticus is built from a Swift front end (Atticus.swift) and a Python "brain" (atticus.py). Speech recognition uses Apple's on-device engine, so audio never leaves the Mac. Requests that need intelligence go through Claude Code's command-line mode (claude -p), which means the app piggybacks on a Claude Pro or Max subscription — no separate API key to paste and no extra bill.
Every voice request is confirmed against Apple's final transcript before any AI tool or learned shortcut runs. Hold mode sends when you release the keys; an optional toggle mode sends on the second press and routes recordings longer than 60 seconds to a review screen instead of sending automatically. Escape cancels listening, and closing the overlay stops the microphone and voice.
Skipping the AI
The central design idea is that most everyday requests do not need a model. Opening apps and sites, moving windows, web searches and music playback run as "instant commands" with no AI involved — the project page claims an app opens in 0.08 seconds. Repeated requests become learned shortcuts: the page cites "Open Calculator and Chess" dropping from 5.5 seconds to 0.4 seconds once learned, again with no AI in the loop.
A background watch feature follows the same principle. Saying "tell me when Codex is done" or "watch this page until it says shipped" sets up a check that runs locally every five seconds, so waiting costs nothing; a notification fires when the condition is met, and coding agents are monitored through their logs rather than the screen.
For asks that do use Claude, the project reports a median cost under $0.002 and six seconds across the last 50 requests ($0.00095 and 5.0 seconds as of October 10, 2026) — roughly 0.14 cents of plan usage per ask, or about 700 asks for $1 at API list rates. A test enforces that the system prompt stays under 10,000 characters, since every request pays for it.
Memory, safety and privacy
Atticus keeps persistent memory: "remember…" statements become saved rules attached to later requests, it learns facts such as your name and favourite apps, and corrections are stored as rules so the same mistake is not repeated. All of it is editable in Settings.
Safety defaults are conservative. A safe mode runs AI-generated commands inside Claude Code's sandbox, restricting writes to a scratch folder. Anything that deletes, sends, buys or changes a file asks for confirmation first, and commands like rm, sudo, disk utilities and shutdown are always blocked. Every exact command is recorded in an ask log.
The app does send one random, locally generated identifier to PostHog once a day so the developer can count users — never names, requests or screen contents — and this can be switched off in settings.
Installing
Requirements are macOS 26 or later (for Apple's new on-device speech), a paid Claude Pro or Max plan, and Claude Code, which the setup window can install and sign into itself. Apple's command line tools are needed via xcode-select, and the optional Peekaboo helper lets Atticus click inside arbitrary apps. Installation is a single curl-pipe-to-bash script that downloads to ~/.atticus, builds ~/Applications/Atticus.app, and configures the app to start at login and self-update. The code is MIT-licensed.
Why it matters
Atticus is a concrete example of where desktop voice assistants seem to be heading: hybrid systems that use a large model for understanding and judgment but route predictable actions through fast, free local paths. That split addresses the two loudest complaints about voice assistants — latency and per-request cost — while the sandboxing, confirmation prompts and ask log show one way to give an agent broad access to a Mac without handing over the keys. For Claude subscribers it is also a preview of subscription-backed tooling: the app itself costs nothing because it reuses a plan you already pay for. The trade-offs are real — it needs the very latest macOS, a paid Claude plan, and a curl-to-bash installer that cautious users will want to read first — but the underlying architecture, where waiting is free and the model is billed only when it actually thinks, is a pattern worth watching.
- #voice-assistant
- #macos
- #open-source
- #claude
- #ai-agents