mcp-patch

Static security scanner for Python MCP server code.

43% of popular MCP servers have shell injection vulnerabilities. No existing tool does AST-level scanning with MCP context awareness. This one does.

Real CVEs this would have caught

• CVE-2025-53967 (Framelink Figma MCP) — shell injection via unsanitized tool parameters
• CVE-2025-6514 (mcp-remote, 437K downloads) — arbitrary command execution via unsanitized tool params

Install

pip install mcp-patch
mcp-patch scan my_server.py

Usage

# Scan a single file
mcp-patch scan server.py

# Scan a directory
mcp-patch scan ./servers/

Example output

Scanning server.py...

  CRITICAL  shell_injection  line 14
  subprocess.run(f"ls {path}", shell=True)
  subprocess.run(shell=True) — tool param 'path' flows to shell
  Fix: Use subprocess.run([cmd, shlex.quote(arg)]) without shell=True

  HIGH      path_traversal   line 28
  open(filename)
  open(filename) — tool param 'filename' used as file path without validation
  Fix: Use (base_dir / Path(filename).name).resolve() and verify result starts with base_dir

Found 2 issues (1 CRITICAL, 1 HIGH) in 1 file.

Checks

Only functions decorated with @tool or @mcp.tool() are scanned. Plain helper functions are ignored.

How it works

Pure stdlib. No network calls. No LLM. Parses your Python source with the ast module, finds @tool decorated functions, collects their parameter names, then walks each function body looking for dangerous call patterns where user-controlled params flow into dangerous sinks.

False positives

This is an MVP scanner — it prefers to over-report rather than miss real vulnerabilities. A path_traversal finding on open(filename) is real even if you have runtime validation elsewhere; the fix is to move validation into the same function so the scanner (and reviewers) can see it.

Development

python -m pytest tests/

No external dependencies. Python 3.9+.