AgentSkin Suite

How do I know this won't delete important data?

Unlike LLM-based summarization, AgentSkin is Deterministic Code. It uses an explicit whitelist strategy. If a key is requested in the signals array, the recursive engine is physically incapable of omitting it. By utilizing aliases, you ensure that even inconsistent nomenclature is mapped correctly to your agent's internal schema. It is as safe as a SQL SELECT statement.

Modern LLMs have 2M+ context windows. Why do I need this?

A context window is a bucket; AgentSkin is a filter. Just because a model can read 2 million tokens doesn't mean it should. "Perceptual Drag" occurs when an LLM allocates attention heads to structural noise (JSON brackets, redundant IDs, ads). By pruning this noise, you free up the model's "IQ" to focus on reasoning. Users typically see a 30-40% increase in reasoning accuracy on complex data sets.

What specific improvements will I see?

• GitHub API: 88.3% token reduction
• CLI output (ls -laR): 99.97% reduction
• Agent replies (Caveman): 65% output token reduction
• Combined: 60-80% of wasted context reclaimed

Why use the local reference server?

Running npx agentskin ensures Self-Sovereign Perception. Your private session cookies, local network data, and API keys never leave your host machine. Perception and pruning happen locally, ensuring absolute privacy and zero-latency execution.

What protections prevent malicious URLs?

The reference implementation includes enterprise-grade security measures:

• SSRF Protection: Blocks private network ranges (127.x, 10.x, 172.16-31.x, 192.168.x, 169.254.x) and IPv6 variants
• Cloud Metadata Blocking: Prevents access to GCP, Azure, and Kubernetes metadata services
• Rate Limiting: 60 requests/minute per client prevents abuse
• Input Validation: Zod schemas validate all tool inputs with type coercion
• URL Sanitization: javascript: and data: URL schemes blocked from HTML extraction
• Resource Limits: 30-second timeout prevents resource exhaustion

All security features are open-source and include 4,695 tests across 274 files for continuous validation.

What's in the AgentSkin Suite?

Three complementary tools that eliminate token waste across the full AI agent cycle:

• AgentSkin SSS — Semantic JSON pruning via MCP. Strips 60-88% of API response noise.
• Tokenjuice — Powers 3 unified MCP tools (reduce, estimate_tokens, apply_json_semantic). Up to 99.97% on large outputs. By Vincent Koc (MIT License).
• Caveman — Output compression via prompt engineering. 65% output token reduction. By Julius Brussee.