[ ⌘K ]

Methodology

How MCP-INDEX discovers, classifies, and analyzes MCP servers for security vulnerabilities.

> PIPELINE LOADING...

Disclaimer: All findings on MCP-INDEX are automated analysis results. They indicate potential vulnerabilities that require manual verification. False positives are possible.

01
DISCOVER
GitHub search 5 queries
>
02
CLASSIFY
Server / Client / Unknown
>
03
FETCH
max 8 files, 2 depth
>
04
ANALYZE
LLM audit prompt
>
05
SCORE
0-100, 4 severity levels
> Daily sweep at 22:00 UTC // Re-analysis on repo update // Avg analysis time: 5s per repo // Avg cost: $0.0018 per repo

MCP-INDEX searches GitHub using five keyword queries:

  • modelcontextprotocol
  • mcp server
  • @modelcontextprotocol/sdk
  • mcpServers
  • claude_desktop_config

Each repo is classified using dynamic file discovery (GitHub tree API):

  • mcp_server — Server SDK imports + tool registration
  • mcp_client — Client patterns (McpClient)
  • unknown — MCP keywords but no implementation

Two-phase analysis using deepseek-v4-pro via CrofAI:

  1. 1.Fetch repo tree via GitHub API, rank by entry point score
  2. 2.Follow imports up to 2 levels deep (max 8 files / 200KB)
  3. 3.Phase 1: LLM determines purpose + threat model
  4. 4.Phase 2: LLM audits with purpose-aware decision tree

Score calculated from LLM findings only:

  • Critical: +40, High: +25, Medium: +15, Low: +5, capped at 100
≥ 70: CRITICAL
≥ 40: HIGH
≥ 20: MEDIUM
< 20: LOW
  • Static analysis only — No runtime verification
  • LLM accuracy — False positives possible
  • GitHub-only — npm, PyPI not covered
  • Rate limited — GitHub API may delay analysis