Atticus is a Mac voice assistant that tries not to call the model
A hold-to-talk macOS assistant that shells out to claude -p for hard requests, then saves the steps so the same request skips the model next time.
Atticus is an open source voice assistant for macOS by SaarthurR. You hold Control-Option-Space, say what you want, and let go. It can open apps, move windows, search the web, control Spotify, read the screen, and draft email replies.
Plenty of projects wire a microphone to a language model. This one is more interesting for how much effort goes into avoiding the model.
How it works
The README describes a two-part design. A Swift file handles the hotkey, speech, overlay, and onboarding. A Python file holds the logic: fast commands, the system prompt, and helpers for windows, web, and screen.
Speech is transcribed with Apple’s on-device recognition, and the README says audio never leaves the Mac and is not saved.
After that there are three tiers. Simple requests, such as opening an app, moving a window, or playing music, are matched locally and never reach a model. The author reports an app opening in about 0.08 seconds.
Anything harder goes to Claude, but not through the API. Atticus runs claude -p, the non-interactive mode of Claude Code, so it rides on a Claude Pro or Max plan you already have and needs no API key.
The third tier is the trick. When a model-driven task succeeds, Atticus saves the steps as a shortcut and replays them next time without the model. The README gives one example that went from 5.5 seconds to 0.4. In the Show HN thread, the author says shortcuts are not continuously re-checked. If one fails, Claude diagnoses the failure and rebuilds it.
There are smaller latency tricks too. Model reasoning can begin during a pause in speech, but its tools are held until the transcript is confirmed. A feature called Watch for me, which waits on a window or a coding agent’s log, polls every five seconds with no model in the loop.
Guardrails and budget
Letting a voice command run shell tools is the obvious risk. According to the README, a safe mode runs model commands inside Claude Code’s sandbox, where they can write only to a scratch folder. Deleting, sending, buying, or changing files asks for confirmation, and rm, sudo, disk tools, and shutdown are always blocked.
The repo also sets itself a cost budget: a median under $0.002 per model-backed request, with $0.00095 reported as measured on October 10. New features are expected to fit under the limit.
One thing to know before installing: the README says a random per-Mac ID is sent to PostHog once a day, which can be turned off. Failure reports are opt-in.
Try it yourself
The repo is MIT licensed. According to the README you need macOS 26 or later, a paid Claude plan with Claude Code, the Xcode command line tools, and Accessibility permission for clicking and typing. The README links a separate Windows version built by someone else.
We have not run this ourselves; this is based on the README and the author’s comments on Hacker News.