Skip to content

SKILL.md

Latest commit

 

History

History
107 lines (77 loc) · 2.94 KB

SKILL.md

File metadata and controls

107 lines (77 loc) · 2.94 KB
name plasmate
description Browse the web using Plasmate for 16x fewer tokens than raw HTML. Use when fetching documentation, reading web pages, extracting structured data, or interacting with websites.
config
mcp
command args
npx
-y
plasmate-mcp

Plasmate Web Browsing

Browse web pages using Plasmate's Semantic Object Model (SOM) instead of raw HTML. SOM compresses web content by 16.6x on average while preserving page structure, element roles, and interactive affordances.

When to use

  • Reading documentation, API references, or READMEs
  • Fetching web page content for research or analysis
  • Extracting structured data from websites
  • Interacting with web pages (clicking, typing, navigating)
  • Any task where you need web content with minimal token usage

Available tools

Once the MCP server is connected, you have these tools:

fetch_page

Fetch any URL and get the Semantic Object Model (SOM). Returns structured JSON with semantic regions (navigation, main, footer), typed elements (links, buttons, headings, tables), and available actions.

fetch_page(url: "https://example.com")

extract_text

Get clean, readable text from any URL. Use this when you only need the text content without structural metadata.

extract_text(url: "https://docs.example.com/api")

open_page

Open a persistent browser session for multi-step interactions.

open_page(url: "https://example.com/login")

click

Click an element by its SOM ID within an open page session.

click(session_id: "...", element_id: "btn-submit")

evaluate

Execute JavaScript in an open page session.

evaluate(session_id: "...", expression: "document.title")

close_page

Close a browser session when done.

close_page(session_id: "...")

Setup

The skill auto-configures the MCP server. If you need to install Plasmate manually:

# Install the MCP wrapper
npm install -g plasmate-mcp

# Install the Plasmate engine (pick one)
brew tap plasmate-labs/plasmate && brew install plasmate
# or: cargo install plasmate
# or: pip install plasmate

Then add the MCP server:

claude mcp add plasmate -- npx plasmate-mcp

Tips

  • Prefer extract_text for simple reading tasks; it returns fewer tokens than fetch_page
  • Use fetch_page when you need to understand page structure or find interactive elements
  • Use open_page + click for multi-step workflows like form filling or navigation
  • Set the PLASMATE_CACHE_API_KEY environment variable for faster cached lookups via the SOM Cache CDN

Why Plasmate instead of raw HTML?

  • 16.6x fewer tokens on average across real websites
  • 50x faster than headless Chrome
  • 30MB memory vs Chrome's 300MB+
  • Structured output with semantic roles, not tag soup

Links