Building AI agents client-side JavaScript

关于

Hey HN, most agent systems default to server-side Python inside containers and chain frameworks. I wanted to see how far we could push agent in the browser with vanilla JavaScript https://buttercup.sh The reason this is interesting is because agent loops in the browser keeps infrastructure costs low. No need for proxy or API calls. And ollama/vLLM can be used for 100% offline. Also WebLLM for embedded. We need to consider CORS, API keys for remote models, access to visual state, and handling remote tool calls. I am working on a guide with references in vanilla JS. This is a short-lived guide starting mid-September with weekly topics. Draft topics starting mid-September: In-Browser Loops: Function calling, and deterministic multi-turn loops running purely in the browser runtime. Vision (Multimodal): Capturing viewport screenshots using browser APIs. Remote Agent Access & Transports: Connecting the in-browser agent to remote agents. Multi-Agent Coordination: Lightweight client-side agent, specialist delegation, and running concurrent sub-agents.

收录记录

1. Show HN

   2026年9月4日 · 1 次收录