⚡ ContextSkeleton

Zero-latency Structural Code Folding CLI, Model Context Protocol (MCP) Server & Token Optimizer for AI Agents (Cursor, Claude Code, Windsurf, AGY).

ContextSkeleton folds function and class implementations while preserving top-level signatures, interface definitions, exported types, and docstrings across JavaScript, TypeScript, Python, Go, and C-style languages.

Token reduction scales with function size and implementation depth — typically 0% on trivial 1-2 line utility functions (comment-marker overhead exceeds the savings), 45-70% on production-sized modules with substantial function bodies, and up to 90%+ on large, implementation-heavy files. Across our internal benchmark set spanning small, medium, and large files in TypeScript, Python, Go, and Rust, the blended average was 67%.

🚀 Quick Start

1. Run CLI Scan

Scan your project to view code compression and token savings metrics:

npx context-skeleton scan

2. Copy Formatted Context for AI Chat Prompts

Dump the compressed skeleton context directly to stdout for Cursor, Claude Code, or Windsurf:

npx context-skeleton copy > context.md

3. Add to PRs & .cursorrules / CLAUDE.md

Inject automated token-savings badges into your repository:

npx context-skeleton badge

Related MCP server: token-pilot

🤖 Model Context Protocol (MCP) Integration

Integrate ContextSkeleton natively with Claude Code, Cursor, or AGY agents.

Add to your claude_desktop_config.json or .cursorrules:

{
  "mcpServers": {
    "context-skeleton": {
      "command": "npx",
      "args": ["-y", "context-skeleton-mcp"]
    }
  }
}

Exposed MCP Tools:

• get_repo_skeleton: Returns folded structural skeleton of target repository.

• unfold_symbol: Retrieves exact implementation of a specific function or class on demand.

• get_token_savings: Returns exact token & prompt cost savings metrics.

📊 Features & Benchmarks

Note: Token reduction scales with function size and implementation depth — typically 0% on trivial 1-2 line utility functions (comment-marker overhead exceeds the savings), 45-70% on production-sized modules with substantial function bodies, and up to 90%+ on large, implementation-heavy files. Across our internal benchmark set spanning small, medium, and large files in TypeScript, Python, Go, and Rust, the blended average was 67%.

💻 Tech Stack ($0 Infra Cost)

• Engine: Zero-dependency Node.js Structural Signature Pruner & Token Counter

• MCP Server: Stdio JSON-RPC 2.0

• Web App: Single-Source HTML5 + Modern CSS + Pure JS

📄 License

MIT License © 2026 ContextSkeleton