AgentSkin Suite

Standard Pruning Example

A typical implementation of the protocol standardizing a verbose meteorological API payload.

The MCP Request

{
  "url": "https://api.weather.gov/gridpoints/TOP/31,80/forecast",
  "signals": ["temperature", "windspeed", "shortforecast"],
  "aliases": {
    "temperature": "temp",
    "shortforecast": "forecast"
  }
}
            

The Resulting Skin

periods[0].temp: 45
periods[0].windspeed: 10 mph
periods[0].forecast: Mostly Clear
periods[1].temp: 38
periods[1].windspeed: 5 mph
periods[1].forecast: Sunny
            

88.3% Token Reduction

GitHub API responses are rich with metadata — node_id, owner.profile, permissions, topics. AgentSkin strips to just the essentials.

Before (1,544 tokens)

{
  "id": 10270250,
  "node_id": "MDEwOlJlcG9zaXRvcnk...",
  "name": "next.js",
  "full_name": "vercel/next.js",
  "owner": {
    "login": "vercel",
    "id": 113386749,
    "avatar_url": "https://...",
    "gravatar_id": "",
    "url": "https://api.github.com/users/vercel",
    ...
  },
  "description": "The React Framework",
  "stargazers_count": 127000,
  "language": "JavaScript",
  ...
}
            

After (180 tokens)

name: next.js
repo: vercel/next.js
description: The React Framework
stars: 127000
language: JavaScript
forks: 26000
open_issues: 2400
license: MIT
updated: 2026-07-09T12:00:00Z
            

99.97% Reduction on Directory Listings

The bash hook transparently compacts noisy CLI output.

Before (3.19M chars)

$ ls -laR agentskin/backend/
agentskin/backend/:
total 551898
drwxr-xr-x 1 user user      4096 Jul  9 12:00 .
drwxr-xr-x 1 user user      4096 Jul  9 12:00 ..
-rw-r--r-- 1 user user      1234 Jul  9 12:00 mcp.js
drwxr-xr-x 1 user user      4096 Jul  9 12:00 lib
drwxr-xr-x 1 user user     65536 Jul  9 12:00 node_modules
...
(56,947 lines total)
            

After (897 chars)

agentskin/backend/:
total 551898
mcp.js (1234 bytes)
lib/ (4 files)
node_modules/ (65536 bytes)
...
[9,185 lines omitted, sha256: abc123...]
            

Web Page Extraction

AgentSkin handles both APIs and web pages. When fetching HTML content, the engine parses semantic structure and converts to structured JSON before pruning.

Processing Pipeline

1. Fetch HTML content
2. Parse with cheerio (HTML parser)
3. Extract: title, h1/h2, paragraphs, links, meta
4. Convert to structured JSON
5. Prune with signals/aliases
6. Output clean Markdown skin