Convert HTML pages into a compact semantic object model for AI agents, reducing token usage by 16x. Built for developers integrating web browsing into LLM applications.
io.github.plasmate-labs/plasmate
Local install
STDIO
No auth required
How models use it and what it is built for.
Convert HTML pages into a compact semantic object model for AI agents, reducing token usage by 16x. Built for developers integrating web browsing into LLM applications.
Local install — runs as a subprocess.
Where to find authoritative docs and source for Plasmate.
Paste any of these into Agent Studio after connecting Plasmate.
Common questions about connecting and running Plasmate.
What does Plasmate do exactly?
Plasmate is a browser engine that converts HTML into a Semantic Object Model (SOM), a compressed representation that captures page structure and content while using 16x fewer tokens than raw HTML. This makes it ideal for AI agents that need to browse and understand web pages efficiently.
How do I install and use Plasmate MCP?
Install via `npx plasmate-mcp@0.4.1`. The server communicates over stdio and integrates with MCP-compatible clients. Once connected, you can send HTML content and receive structured semantic representations.
What's the token savings compared to raw HTML?
Plasmate reduces token usage by approximately 16x by converting verbose HTML markup into a lean semantic object model. This significantly lowers costs and latency when processing web pages in AI agent workflows.
Can I use Plasmate for real-time web scraping?
Plasmate is designed as a semantic parser for HTML input, not a live browser. You'll need to fetch HTML separately (via curl, a web client, or another tool) and then pass it to Plasmate for conversion to semantic objects.
What kind of output does the Semantic Object Model provide?
The SOM extracts structured data from HTML—such as text content, navigation, forms, and page hierarchy—in a compact format. Refer to the Plasmate documentation for the exact schema and field definitions supported by the current version.
MCP Playground runs 10,000+ hosted MCP servers — GitHub, Linear, Notion, Stripe, Sentry and more — across Claude, GPT, Gemini, DeepSeek and 60+ AI models. Compare model answers side-by-side, save agent presets, share runs. Zero install.
Open Agent StudioJudges Panel
45 judges that evaluate AI-generated code for security, cost, and quality with built-in AST.
Addozhang Nexus
Query Sonatype Nexus Repository (OSS/Pro) for Maven, Python, and Docker artifacts
Cyanheads MCP Ts Template
TypeScript template for building MCP servers with declarative tooling, observability, and auth.
Chernistry Bernstein
Declarative agent orchestration for engineering teams