AgentSkin

Context middleware for AI agents. AgentSkin removes low-value API, JSON, and terminal noise before it enters model context.

What it does

AgentSkin has three primary workflows:

• compress — one front door that auto-detects JSON, CLI output, or plain text.
• fetch_optimized_data — fetches a public URL and applies URL-specific pruning rules when available.
• reduce — compacts terminal output using Tokenjuice's rule-driven reducers.

Five advanced tools remain for compatibility and diagnostics: apply_json_semantic, classify_url, strip_ansi, estimate_tokens, and skin_reasoning.

Quick start

npx -y agentskin@latest

MCP config:

{
  "mcpServers": {
    "agentskin": {
      "command": "npx",
      "args": ["-y", "agentskin@latest"]
    }
  }
}

Design rule

Fidelity first, compression second. Explicit signals and URL rules are authoritative. Generic keys such as id, name, and url are only fallback signals when no explicit rule exists.

A reduction is useful only when the information required by the task survives it.

Current verification

The release gate runs the AgentSkin-owned test suite, including MCP startup/security and live HTTP integration tests. Run it with:

npm ci
npm test

Benchmarks are workload-specific. The test suite prints measured reductions for its fixtures instead of claiming one universal savings percentage.

Runtime

• Node.js MCP server over stdio
• Tokenjuice is a normal npm dependency for CLI reduction
• URL-specific semantic pruning rules for structured APIs
• SSRF checks for fetches, response-size limits, rate limiting, Zod validation, and processing timeouts

Development

This directory is the npm package root.

npm ci
npm test
npm run lint
npm pack --dry-run

Credits

AgentSkin semantic pruning and MCP integration: Shawn Nichols Sr. / Nichols Transco LLC.

CLI reduction uses Tokenjuice by Vincent Koc under its MIT license.