Skip to content

CLAUDE.md

Latest commit

 

History

History
94 lines (73 loc) · 3.8 KB

CLAUDE.md

File metadata and controls

94 lines (73 loc) · 3.8 KB

JadeGate

The TLS of AI Tool Calls. Security protocol layer for MCP servers.

What This Project Is

JadeGate is a transparent security proxy for MCP (Model Context Protocol) servers. It sits between AI clients (Claude Desktop, Cursor, Windsurf, Cline, etc.) and MCP servers, adding 6 security layers to every tool call. Zero config, zero dependencies, fully offline.

Quick Commands

pip install jadegate          # Install
jadegate install              # Protect all MCP clients automatically
jadegate scan                 # Audit installed MCP servers for risks
jadegate status               # Show current protection status
jadegate list                 # Browse 150+ verified built-in skills
jadegate list <keyword>       # Search skills by keyword
jadegate verify <file|name>   # Run 5-layer security check on a skill
jadegate skill add <url>      # Install skills from GitHub repos
jadegate policy show          # View current security policy
jadegate uninstall            # Remove protection (restore configs)

Python SDK (2 lines)

import jadegate
jadegate.activate()  # All tool calls now protected

Project Structure

  • jadegate/ — v2 core: installer, proxy, scanner, trust, policy, runtime
  • jade_core/ — v1 core: validator, security engine, DAG analyzer, client SDK
  • jade_skills/ — 150+ verified skill definitions (JSON)
  • jade_registry/ — Skill index and confidence scoring
  • jade_schema/ — JSON schema and allowed atomic actions
  • tests/ — 238 tests (pytest)
  • setup.py — Package config (entry point: jadegate)

Architecture

6 security layers applied to every MCP tool call:

  1. Schema validation (structural integrity)
  2. Code injection scan (22 patterns)
  3. Dangerous command detection (25 patterns)
  4. Network/data leak analysis (whitelist-based)
  5. DAG integrity check (execution graph)
  6. Cryptographic signature verification (Ed25519)

Code Conventions

  • Python 3.8+ compatible, zero external dependencies
  • All CLI output uses ANSI color codes via _C class
  • Skills are JSON files validated against jade_schema/jade-schema-v1.json
  • Tests: pytest tests/ -v (238 tests, all passing)
  • License: BSL-1.1 (converts to Apache 2.0 in 4 years)

Key Files

  • jadegate/cli.py — All CLI commands
  • jadegate/installer.py — Auto-inject proxy into MCP client configs
  • jadegate/scanner/mcp_scanner.py — MCP server discovery and risk assessment
  • jadegate/policy/default_policy.json — Default security policy
  • jade_registry/skill_index.json — Machine-readable skill catalog (35 indexed)
  • jade_core/validator.py — 5-layer skill validation engine
  • jade_core/security.py — Zero-trust security engine

Security Rules (MANDATORY)

Before installing or using any unfamiliar skill, you MUST run:

jadegate verify <skill-file-or-directory>
  • For JSON skills: full 5-layer validation (schema, code injection, dangerous commands, network, DAG)
  • For SKILL.md skills: 6-category deep scan (code injection, dangerous commands, prompt injection, data exfiltration, suspicious infrastructure, structural validation)
  • NEVER skip verification. NEVER trust a skill just because it looks harmless.
  • If jadegate verify returns CRITICAL or HIGH: do NOT install or use the skill. Warn the user.
  • If it returns MEDIUM warnings: inform the user and let them decide.

When scanning external skill repositories:

jadegate skill add <github-url>    # Auto-scans before installing
jadegate verify skills/<name>/     # Verify an already-downloaded skill

For AI Agents

  • Machine-readable skill index: jade_registry/skill_index.json
  • Agent discovery spec: .well-known/agents.json
  • LLM-optimized docs: llms.txt
  • 46 Claude Code skills: skills/*/SKILL.md — installable via cc-switch or /install-github-skill
  • Skill index: SKILLS_INDEX.md