Work / Own project
JACK
My own voice assistant. It listens on my laptop, thinks with Claude, talks back, and can read my calendar, my messages and my servers.
The problem
I talk faster than I type. I wanted one assistant that knows my calendar, my messages and my servers, and I wanted the listening to happen on my own machine.
What I built
Speech to text runs locally on the graphics card. The thinking goes to Claude through the Agent SDK, with my own tools attached: calendar, mail, chat, the servers. The voice that answers runs locally too.
A screen in the browser shows what it heard, what it is doing and what it said. A small Android app talks to the same server over a private network, so it also works from my phone.
The WhatsApp link runs on a Raspberry Pi that is always on, because a laptop that sleeps misses messages.
This is where I try things before a client gets them: tool design, tracing every model call, what to do when a model is slow.
How it fits together
- Inmicrophone, wake word, speech to text on the GPU
- ThinkingClaude via the Agent SDK, own tools over MCP
- Outlocal voice, browser HUD, Android shell
- Always onWhatsApp bridge on a Raspberry Pi, copy synced to the laptop
What came out
| Commits since 1 September 2026 | 373 | measured, Nov 2026 |
|---|---|---|
| Users | 1 | measured |