MCP-scan – Security scanner for MCP server configs
First security scanner for MCP configs as the protocol gains adoption.

It turns real program analysis — ASTs, cross-file dependency graphs, taint tracking and Z3 symbolic paths — into callable MCP tools for agents, not just another prettier linter. Concrete features like simulate_refactor, generate_unit_tests from symbolic paths, and cross-file security_scan give it a distinct technical voice. The <10% false-positive claim and heavy test coverage are promising, but I'd want to see results on large, messy repos before swapping out existing scanners.
Developers and security engineers using AI agents or CI to analyze and refactor codebases
Security detection: - 16+ vulnerability types (SQL/NoSQL/LDAP injection, XSS, command injection, SSRF, CSRF, SSTI, prototype pollution, weak crypto, etc.) - Taint tracking across files - <10% false positive rate - Z3 symbolic execution for path analysis
The MCP part exposes 23 tools that let AI agents analyze code: - security_scan, cross_file_security_scan - symbolic_execute (Z3-based path exploration) - generate_unit_tests (test gen from symbolic paths) - simulate_refactor (behavior preservation check) - code_policy_check (compliance verification)
CLI works standalone too. No-install usage: uvx codescalpel mcp
Or: pip install codescalpel
Testing was important - 7,297 test cases with 94.86% coverage.
What I'm curious about: - Is <10% false positive rate good enough for AppSec teams? - What other security checks would help? - Interest in expanding to Go/Rust/C++?
Target users: Individual developers (cost reduction story), security engineers (OWASP Top 10 evaluation), team leads (ROI analytics), enterprise architects (SOC2/ISO compliance).
Repo: https://github.com/3D-Tech-Solutions/code-scalpel
MIT licensed, actively maintained. Feedback welcome!
First security scanner for MCP configs as the protocol gains adoption.
Tree-sitter AST compression cuts LLM context tokens 50-70% while preserving API structure.
Forces LLMs to debug with AST evidence instead of pattern-matching symptoms.
Catches AI code debt deterministically with AST analysis when linters only check style.
Audit report without the actual tool repo — methodology is interesting, but where's The Janitor?
AST + embeddings for codebase search—but Sourcegraph Cody, Cursor, and Continue already solve this.