From a single CLI tool to 60+ repos across LangChain, Zapier, Supabase, VS Code, and more. Here's how the Plasmate ecosystem grew and what it means for AI agent developers.

We benchmarked 100 real websites and found Plasmate's Semantic Object Model achieves 30x mean token compression over raw HTML. Top performers like accounts.google.com hit 864x. Here's the data.

A practical guide to building an AI agent that aggregates news from 50 sources, extracts headlines and summaries via SOM, and produces a daily briefing. With daemon mode, the entire pipeline runs in under 60 seconds.

The biggest complaint about Plasmate was cold-start latency. Daemon mode keeps the browser warm between fetches, cutting per-page time from 2 to 5 seconds down to 200 milliseconds.

The Plasmate SOM compiler is now available as a 380KB WebAssembly package. Compile HTML to structured JSON in Node.js, Deno, Bun, browsers, serverless functions, or edge workers. No native binary required.

Publishers can now feed HTML directly to the Plasmate compiler from files or stdin. No browser, no network requests, no extra infrastructure. Just HTML in, SOM out.

We benchmarked three web representations across 100 tasks and two models. SOM cuts tokens 4x vs HTML and is faster than markdown on Claude.

Build a Python agent that monitors competitor product pages, extracts pricing data via SOM, detects changes, and sends alerts. Complete working code included.

v0.5.0 adds html_id for DOM resolution, ARIA state preservation, details/summary extraction, improved Wikipedia tables, GDPR banner stripping, and ICU/Intl support for JS-heavy SPAs.

Instead of blocking AI agents or letting them scrape raw HTML, publishers can serve structured SOM representations directly. A step by step implementation guide.

Your AI agent needs to browse sites you are logged into. Twitter, LinkedIn, internal tools. Here is how Plasmate handles authenticated sessions without sharing passwords.

A complete tutorial for building a web research agent that uses SOM to browse the web at 4x lower token cost. Python, OpenAI, and Plasmate. Full code included.

The web has served humans, then search engines, then applications. AI agents are the fourth consumer, and they have no format designed for them. We propose three infrastructure primitives to fix this.

HTML was designed for browsers. Markdown was designed for humans. Neither was designed for AI agents. SOM is the missing format for the fourth consumer of the web.

Raw HTML wastes 80% of your LLM context window on CSS classes, tracking scripts, and layout noise. Plasmate is an open-source headless browser that compiles web pages into structured semantic JSON for AI agents.