License Apache-2.0 v0.5.1 Rust

The browser engine for agents.

Plasmate compiles HTML into a Semantic Object Model, a structured representation that LLMs can reason about directly. 50x faster than Chrome. 10x fewer tokens.

1,000 pages/day: $99 in HTML tokens vs $5.69 in SOM. Same content. 94% less cost.

Reduced motion is enabled. Background particles are disabled.

Speed
4-5 ms/page

Chrome takes 252ms. Lightpanda takes 23ms. Plasmate takes 4.

Memory
28 MB

Chrome uses ~20GB for 100 concurrent pages. Plasmate uses 28MB.

Tokens
94% token savings

Across 98 sites. Up to 1,544x on heavy SPAs. Mean 44.9x, median 9x. See the benchmark.

Cost
$5.69 vs $99

Per 1,000 pages/day with GPT-4o. SOM tokens vs raw HTML tokens. Calculate yours.

The DOM was built for humans

The Document Object Model represents a page as a tree of elements. It includes styling, layout, accessibility attributes, event handlers, and rendering state. All of it.

Agents don't need all of it.

The Semantic Object Model extracts what matters: structure, content, and interactive elements. It's a projection of the DOM optimized for reasoning, not rendering.

DOM

47,000 tokens

SOM

4,500 tokens

Wikipedia homepage, same page.

SOM is not lossy compression. It's a different representation for a different purpose.

7 methods. That's the protocol.

Chrome DevTools Protocol has hundreds of methods. Most exist for debugging, profiling, and human-driven inspection.

AWP has seven:

navigate snapshot click type scroll select extract

These are intents, not instructions. Plasmate handles the implementation. Your agent handles the reasoning.

If you've used Puppeteer, the API will feel familiar. The token cost will not.

CDP compatibility is included as a legacy bridge. AWP is the native protocol.

Infrastructure, not product

Plasmate is Apache 2.0 licensed. Use it, fork it, embed it, sell it.

SOM and AWP are specifications, not features. We want them implemented everywhere: in other browser engines, in agent frameworks, in production systems we'll never see.

The web has standards for documents, for styling, for scripting. It needs a standard for agents. This is that standard.

Member of the W3C Web Content and Browser AI Community Group. SOM and AWP are being developed as open web standards.

30 seconds. Your AI gets a better browser.

Claude Code

One command. No config files.

claude mcp add plasmate -- npx plasmate-mcp

OpenClaw

One command. Skill handles everything.

clawhub install plasmate

Cursor / Claude Desktop

Add to your MCP config, restart.

npx plasmate-mcp

CLI / Binary

For scripts, pipelines, and direct use.

curl -fsSL https://plasmate.app/install.sh | sh

First browser tool on the official MCP Registry. 230 tests passing. 94/98 top sites covered.

No install required

Don't want to run a binary? The SOM Cache is a hosted API that compiles any URL to SOM on demand. Shared CDN, cached results, zero infrastructure on your side.

When multiple agents request the same page, they share a single compilation. Less redundant fetching for publishers, lower latency for agents.

$ curl https://cache.plasmate.app/v1/som?url=https://example.com

Works with any HTTP client, any language, any agent framework. No API key required for public pages.

Serve SOM alongside HTML

Publishers can pre-compute SOM for their pages and serve it at a well-known endpoint. Agents discover it automatically, just like sitemaps for search engines.

/.well-known/som.json

This is the cooperative path: instead of blocking agents or serving them expensive raw HTML, give them a structured format designed for their consumption. Lower bandwidth costs, content control, and provenance metadata for attribution.

The economics: mid-size publishers serving 1M agent requests/month save ~68% on infrastructure by serving SOM instead of HTML. Read the full analysis.

plasmate.app and somordom.com already serve /.well-known/som.json. robots.txt extension proposal in the research papers.

What's new

SPA rendering bridge with bidirectional V8/DOM sync. Network interception. TLS fingerprint tuning. Wasm plugin system. Multi-page sessions (50 concurrent). page.click(), page.type(), page.waitForSelector(). Auth profiles with encrypted storage. ~25K lines of Rust. 230 tests passing.

Performance in context

Engine Per page 100 pages Memory Binary
Plasmate 4-5 ms 405-512 ms 28 MB 45.8 MB
Lightpanda 23 ms 2,300 ms ~2.4 GB 59-111 MB
Chrome 252 ms 25,200 ms ~20 GB 300-500 MB

Local harness, 100 pages, 231KB average HTML. No caching. Methodology matches Lightpanda's published benchmarks.