Hi from Rennes: running JS LLM agents inside Swift apps

Hi all, I’m Sébastien, based in Rennes, France. I’ve been building software for 30+ years. That includes being a developer and partner at Kinoma in California, where I first worked with the XS JavaScript engine. Today I run Haruni, a small applied-ML studio.

What I’m building

• KaozKit: a framework for running JavaScript LLM agents inside native Swift apps, on top of the XS engine from Moddable. The agent logic lives in JS, and the host app decides which tools and data the agent can reach.
• TyKaoz: a privacy-first native macOS LLM client built on that stack, designed local-first.

What’s working
Local inference is now good enough for real agent workloads. I run vLLM on a DGX Spark and MLX on Apple silicon. Keeping agent logic in a small embedded JS engine has made it much easier to iterate without rebuilding the app.

One thing I need
Feedback from Swift/macOS developers who ship LLM features: how do you structure agent logic today, and what would it take for you to embed a JS runtime for it? I’m also happy to meet anyone here in Rennes.