Image: GitHub
Why it mattersA coding agent that runs long tasks on a developer's Mac does not need the developer at the desk to approve the next command, which is the step that turns an agent run into a thirty-minute babysit.
A hold-to-talk Apple Watch app that reaches the AI coding agent running on your Mac, so you can push a project forward from across the room or from a walk, is the shape of Chums. The project went up on GitHub and has 90 stars after five days under an MIT licence. The idea is that the AI subscription stays what it is, and the watch and iPhone replace the terminal window as the place a developer answers from.
What the watch does
The watch app draws a character for each agent, and the character acts out what the agent is doing: listening, thinking, working, waiting for the user, done, blocked. Holding the screen records speech, and letting go sends what was said to the agent through the Mac. When the agent asks for permission to run a command, a card appears and the user holds to allow or taps to deny. Short answers are spoken on the watch in the agent's own voice, and the full conversation scrolls. A keyring mode keeps the app on screen with the watch off the wrist, so a long run can be followed on a desk. Each agent is a separate session with its own character, and switching between them is a tap on the faces at the top left.
How it reaches the Mac
The watch and iPhone cannot reach an address on a home network, so Chums publishes a hub on the Mac through a Tailscale Funnel, protected by a token the watch app holds. The hub, written in Node.js, holds the agent sessions, routes voice, transcribes speech with WhisperKit, and voices replies with Kokoro, all on the Mac. The iPhone app shows the same conversations with room to type, send photos and files, and open live outputs: an HTML chart the agent makes opens as a page, Word, Excel and Keynote files open in Quick Look.
What the setup takes
Chums is source code, not an App Store download. The project targets an Apple Watch on watchOS 26, the paired iPhone on iOS 26, a Mac on Apple silicon with Xcode 26, an Apple ID, a free Tailscale account, and an existing AI coding agent. The first setup takes 20 to 30 minutes, plus up to half an hour of waiting for a new Funnel address to reach public DNS. The README gives a one-line prompt that runs the setup and stops at the steps a human has to perform: signing in, plugging in the iPhone, and turning on Developer Mode.
Where this fits
A coding agent run that lasts an hour is spent half at the keyboard and half approving the next command. A developer who steps away from the Mac can still see the character working and give the next approval from the watch or the iPhone, which turns a long run from a babysit into a background job. The AI subscription stays with the user. The hub holds no account of its own, and the Mac does the speech work on device, which is the quiet part of the pitch for anyone who does not want speech recordings leaving the laptop.
Source
The repository is on GitHub at lcoutodemos/chums. The project site is at chums.club.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.