AgentSkin Suite

Get Started in 30 Seconds

Run the unified MCP server with a single command:

npx -y agentskin@latest

Then add to your agent's MCP config:

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

That's it. You now have 7 tools: API pruning, CLI reduction, token estimation, and more.

The Complete Token Optimization Stack

AgentSkin Suite combines 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 — Rule-driven CLI compaction + powers 3 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.

Combined, they reclaim 60-80% of context that was previously wasted on noise.

One Unified MCP Server, 7 Tools

The Suite exposes 7 tools through a single MCP server — AgentSkin's semantic pruning merged with Tokenjuice's reduction engine (by Vincent Koc, MIT License):

• fetch_optimized_data — Fetch URL, prune JSON, return Markdown skin (88% savings)
• skin_reasoning — Strip linguistic noise from text (14-29% reduction)
• classify_url — Match URL against 11 built-in API rules
• strip_ansi — Strip 5 patterns of ANSI escape codes
• reduce — Tokenjuice reduction pipeline (up to 99.97%)
• estimate_tokens — Grapheme-aware token estimation
• apply_json_semantic — Prune raw JSON with signal keys

echo "source $(pwd)/.agents/hooks/bash-optimizer.sh" >> ~/.bashrc && source ~/.bashrc

The bash hook intercepts noisy CLI commands and pipes output through Tokenjuice before the agent sees it — zero overhead, zero agent awareness.

Enterprise-Grade Protection

The reference implementation includes robust security measures to protect against common attack vectors in AI agent pipelines.

SSRF Protection

Blocks requests to private network ranges including IPv4 (127.x.x.x, 10.x.x.x, 172.16-31.x.x, 192.168.x.x, 169.254.x.x) and IPv6 variants (::1, ::ffff:, fe80:). Zone IDs are stripped and validated before processing.

Cloud Metadata Blocking

Prevents access to cloud provider metadata services (metadata.google.internal, metadata.azure.com, kubernetes.default.svc) that could expose cloud credentials or configuration.

Rate Limiting

60 requests/minute sliding window per client to prevent abuse and ensure fair resource allocation in multi-agent environments.

Input Validation

All tool inputs are validated with Zod schema validation. Type coercion ensures signals, aliases, and apply_reasoning parameters are properly typed before processing.

URL Sanitization

HTML-extracted URLs are sanitized to block dangerous schemes (javascript:, data:) that could enable XSS attacks through the MCP tool interface.

Resource Limits

30-second processing timeout prevents resource exhaustion from maliciously large payloads.

Open Source Auditing

The security implementation is fully open-source and includes 4,695 tests across 274 files for continuous validation.

Performance Metrics

The reference implementation delivers verifiable, deterministic compression results.

Token Reduction

• GitHub API: 88.3% (1,544 → 180 tokens)
• CLI ls -laR: 99.97% (3.19M → 897 chars)
• Caveman output compression: 65% average
• Caveman memory file compression: 46% smaller
• Combined test suite: 4,695 tests, 274 files — 100% passing
• Pipeline throughput: 3,030 fixtures/sec (0.33ms avg)

Processing Time

Sub-100ms transformation latency for payloads under 100KB. 30-second maximum timeout for large payloads.

Unified 7-Tool MCP Server

v5.0 merges AgentSkin's semantic pruning with Tokenjuice's reduction pipeline into a single MCP server:

• Unified server — One npx agentskin@latest instead of two separate servers
• 3 new tools — reduce, estimate_tokens, apply_json_semantic from Tokenjuice
• Tokenjuice integration — 136 rules, 143 fixtures, 3,030 fixtures/sec pipeline
• Rate limit unified — 60 req/min sliding window (was 30/60 split)
• Backward compatible — All existing AgentSkin tools unchanged

Special thanks to Vincent Koc for the MIT-licensed Tokenjuice engine that powers the reduction pipeline.

Built By

Creator	Contribution
Shawn Nichols Sr. Nichols Transco LLC	AgentSkin SSS protocol, MCP server, Suite integration
Vincent Koc github.com/vincentkoc/tokenjuice	Tokenjuice — MIT-licensed engine: powers reduce, estimate_tokens, and apply_json_semantic in unified MCP server
Julius Brussee github.com/JuliusBrussee/caveman	Caveman — Output compression via prompt engineering