GitHub Gems

Atticus Mac Voice Assistant Runs on Your Claude Plan, No API Key

Atticus is a free, open-source voice assistant for Mac that runs on a paid Claude plan without an API key. Its repository on GitHub describes fast local commands and memory.

A Mac laptop screen showing a voice assistant overlay with a microphone icon above an open application window
Photo: GitHub

A new open-source project called Atticus is drawing attention among Mac users who want a voice assistant that does not require a separate API key. According to its repository on GitHub, the tool lets people hold the shortcut ⌃⌥Space, speak a request, and release the keys to have it carried out. It can move windows, find files, read the screen and learn repeated tasks. The project says it runs on an existing paid Claude Pro or Max subscription.

What the project does

The repository describes Atticus as a voice assistant that works across the whole Mac. Users can ask it to arrange windows, locate documents and respond to what is currently on screen. The project page highlights four traits: speed, no API key requirement, memory of the user, and a free, open-source licence.

GitHub's project summary says simple requests skip the AI entirely. In the example given, opening an app takes 0.08 seconds. When a user asks for a particular action once, Atticus records it as a learned shortcut, so the same request can run again without AI and at no cost.

Memory is the other main feature. Saying 'remember' before a statement stores it for future requests. According to the page, Atticus picks up details such as the user's name and frequently used apps, and it can correct its own mistakes when the user points them out.

How it is controlled

Users can choose between two modes in Settings. In Hold mode, speech is sent when the shortcut is released. In Toggle mode, the first press starts listening and the second press sends the request. The repository notes that pressing Esc cancels listening without sending anything, and that toggle recordings open a review screen after 60 seconds instead of sending automatically.

Voice input is handled by Apple's on-device speech system, which the project says requires macOS 26 or later. Before any AI tool or learned shortcut runs, Atticus waits for Apple's final transcript to confirm the request. That step is a safeguard: a partly heard sentence should not trigger an action on the computer.

Visual feedback and privacy

While Atticus works in a window that is covered, on another desktop space, or on another display, a small live preview shows that window. The repository states that this preview stays on the Mac, is never sent to the AI and is not saved. Without Screen Recording permission, the card shows only the app icon and window title.

Meeting Copilot and local archives

The project also includes a Meeting Copilot feature, which must be enabled in Settings. The repository says meeting transcripts are saved by default, with speaker labels, timestamps, notes and answers written to a folder on the user's Mac. Those folders are set to private permissions, and the page says audio is never saved.

Meetings can be searched, exported as Markdown or deleted from the local archive. If AI meeting summaries are switched on, which is the default, Claude reads the transcript once after a session ends and writes a summary and action items under the user's own Claude plan. Users who want transcripts to stay entirely on their Mac can turn that setting off, and the tool then produces a simpler local recap instead.

Feedback and telemetry

Users can send feedback or report a bug through the menu bar or Settings. The repository says the draft shows its exact contents before it is opened as a GitHub issue, and that answers, screen content and command logs are never attached. The page also discloses that once a day Atticus sends a random identifier generated on the user's Mac to PostHog to count users. It says the identifier is not tied to the person, and that the count can be turned off in Settings.

Cost and installation

The project publishes a budget target for AI requests. It lists a median cost under $0.002 per AI ask and a time limit of six seconds over the last 50 asks. As of October 10, 2026, the repository reports a median of $0.00095 and 5.0 seconds. Setup requires macOS 26 or later, a paid Claude plan and Claude Code, plus Apple's command line tools and Accessibility permission. The installer places the app in the Applications folder and sets it to start at login.

For readers considering it, the main trade-off is clear. Atticus relies on a subscription they already pay for rather than metered API billing, and on Apple's speech system rather than a third-party service. It is also a project that asks for broad control of the computer, so the permission requests and safety checks described in the repository deserve attention before installation.

The project is released under the MIT licence, and its repository invites contributions. Developers who want to check the work can run the test suite described in the documentation, but general users should expect to rely on the project's own description of its behaviour until independent reviews appear.

Source: GitHub

Comments (0)