According to the Lookma README, users hold the fn key to speak a request in any language, tap fn for a hands-free conversation, or type through the menu bar. The voice layer passes those words to a local core on 127.0.0.1, which prompts Claude Code or Codex. The agent interacts with real Mac apps through seven hands tools, showing a coral ring and ghost cursor on each target before acting while waiting whenever the user is typing.1
How to install and run it
The README documents installing Lookma by cloning the GitHub repository, running npm install and npm link, checking setup with lookma doctor, and starting a session with lookma ~/Projects. Running it requires macOS 15 or later, Node.js 20 or later, Xcode Command Line Tools to compile Lookma.app with swiftc, and either Claude Code 2.1.330 or later or Codex CLI 0.163.0 or later. First launch requests Accessibility, Screen Recording, and Microphone permissions.1
What it produces and how approvals work
In practice, a user can ask Lookma to turn on dark mode or find three ramen places near the ferry and save them in a new note, according to the README. The app shows an approval card the first time it controls any app, storing grants in ~/.lookma/apps.json, and surfaces agent permission prompts as cards. The README reports that the voice model alone approved a lone approve-question 15 of 15 times, a go-ahead said 0.3 seconds after the card 11 of 16 times, and its own leaked voice 4 of 12 times.1
Stated limits and safety boundaries
The README lists several constraints: Lookma supports only the main display on macOS, runs one agent session at a time, and cannot drag from one app to another. A task of a few steps takes 10 to 20 seconds, mostly from model thinking, and voice approval understands only English and Chinese. On safety, the README states that asking before irreversible actions like sending, posting, paying, or deleting is only a model instruction rather than an enforced gate.1