- Agentic web browsing can’t scale if every page action triggers an LLM.

- Small language models can power most interactive AI experiences.

- The browser is the ideal runtime being open, GPU-accelerated, and everywhere.

- Agentary.js is an open-source SDK for building agentic workflows that run locally, in the browser.

Repo URL: https://github.com/agentary-ai/agentary-js

---

Most agentic browsing today sends every click, scroll, or query to a cloud LLM. This doesn’t scale, economically or technically. Each DOM action becomes another API call that is slow and expensive, not to mention a compromise to user privacy.

Agentic behavior belongs inside the browser, not round-tripped to the cloud.

NVIDIA Research even published a paper (https://arxiv.org/pdf/2506.02153) arguing that most agentic tasks don’t need generalist LLMs. They need fast, specialized SLMs running close to the user.

I wanted to build an open-source SDK to support a future where agentic workflows are not only built and owned by the website, but powered by SLMs. With the expansion of WebGPU, the browser is now the perfect runtime.

I put together a simple demo that uses Agentary.js, so feel free to check it out and share feedback: https://demo.agentary.ai

I'd love to collaborate with other engineers in the community to make this future a reality, and importantly open-source!