I built webmodels.dev to explore a simple idea: open-weight AI models should be a shared, permissioned capability of the web.

The proposed navigator.models API would let an app request the exact model and release it was built for. The browser would handle user consent, shared model weights, isolated execution, and device resource limits—making private, on-device AI more practical without every website downloading its own copy of the model.

The site turns the proposal into an approachable visual explainer, with API examples for generation, streaming, and embeddings, plus use cases such as document search, audio transcription, and receipt scanning.

This is an early proposal, not a shipped browser API. I launched the site to invite developers and browser implementers to explore the idea, challenge its assumptions, and help shape the discussion.

If you are at all interested in this idea I would love your feedback or a thumbs up on the proposal: https://github.com/webmachinelearning/proposals/issues/22