PARAS MADAN
All writing

Latest in AI

Atticus: A Mac Voice Assistant with Local Commands and Claude Code

Atticus combines on-device speech, reusable desktop actions and Claude Code, with local monitoring that avoids repeated AI calls while waiting.

Atticus is an open-source Mac voice assistant that handles simple desktop commands locally and sends more complex requests through Claude Code. Familiar actions can be replayed, and background checks can run without repeatedly consulting a model.

The creator introduced it in a Show HN post on October 12. It is an early, source-built project for macOS 26 or later. This overview draws on the documentation and code; it is not a hands-on reliability test.

Atticus uses local commands before calling Claude

The basic interaction is hold Control, Option and Space, speak, then release. Apple's on-device speech recognition turns the audio into text. A toggle-to-talk mode is also available, and requests can be edited in the overlay.

Opening an app, moving a window, searching the web or controlling supported music actions can skip the model. More complex requests go through Claude Code's noninteractive mode, with AI tools waiting for transcript confirmation.

Supported built-in commands can run mid-sentence when the optional instant-action settings are enabled. Learned shortcuts wait for confirmed speech. Check those settings before speaking around open work.

Atticus also saves suitable actions for later replay. The Python implementation restricts this to eligible commands and matches a normalized version of the request. It avoids storing certain context-dependent actions, and a failed replay falls back to the AI path.

Opening a familiar group of apps is a sensible use case for replay. Check saved actions again when the desktop or the intended routine changes.

Waiting does not need a model call every few seconds

A user can ask Atticus to watch for a page to show particular text or notify them when a coding agent finishes. AI sets up the check, then local code handles the waiting.

The implementation checks every five seconds and requires two consecutive matches before triggering. It stops after three hours. Window-based checks need the window to remain open on the current desktop; covered windows are supported, but minimized ones are not. Supported coding-agent checks can use their logs instead of the screen.

When the condition is met, Atticus sends a notification and can run a requested follow-up. Setup and follow-up may use AI; the intervening polling does not.

That design fits bounded tasks such as waiting for a build or an agent run. Conditions requiring fresh judgment on every check need a different approach.

Availability, privacy and permissions matter

Atticus itself is MIT-licensed and free. Its documented setup requires macOS 26, Apple's command-line tools, and Claude Code with a paid Claude Pro or Max plan. Model requests use that plan; a free application does not make its cloud inference unlimited. Optional Peekaboo integration adds the ability to click inside applications.

There are no packaged GitHub releases at the time of writing. The supplied installer downloads and builds the app. Setup also enables login startup and automatic updates, which the documentation says can be disabled. Review the installer and permissions before trying it on a machine containing sensitive work.

The documentation says speech recognition keeps audio on the Mac. Complex requests use cloud-backed Claude, so relevant text or screen context can leave the device.

The project documents Safe mode, command logging and approval prompts for consequential actions. Its code uses Claude's permission-bypass mode alongside its own approval hooks and sandbox settings. Those controls need scrutiny; the documentation alone does not establish that every risky action is caught.

For an initial trial, stay with reversible app and window tasks. Check whether voice input and replay save effort in a routine you repeat before giving the assistant more responsibility.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help developer-tool teams explain how their products work, turn technical tradeoffs into useful launch stories and reach the founders and engineers who can use them.

Book a 30-minute call.