guardrail-mcp

---

GuardRail exposes tools that scan source and infrastructure text for high-signal issues (secrets, dangerous APIs, injection patterns, IaC misconfigurations), with optional tree-sitter structural checks, dependency inventory/OSV, SARIF/SBOM export, and an enterprise policy gateway.

Resource	URL
Website	https://secret4422.github.io/guardrail-mcp/
AlphaTech Studio	https://alphatech.studio
MCPize listing	https://mcpize.com/mcp/guardrail
Accuracy policy	docs/ACCURACY.md
Security / threat model	docs/SECURITY_MODEL.md
Examples	docs/EXAMPLES.md
Performance	docs/PERFORMANCE.md
Test results	docs/test-results.md
Enterprise	docs/ENTERPRISE.md

Scope (read this)

• Does: static analysis of text you provide; redacts many secret-shaped substrings in excerpts.
• Does not: execute scanned code; replace commercial SAST/DAST; guarantee zero false positives/negatives; provide SOC2 certification by itself.
• Website playground: browser-only demonstration (docs/ACCURACY.md). Production use = Python MCP/CLI.

Quick start

Option 1: Fast run with uvx (zero installation)

uvx --from git+https://github.com/SECRET4422/guardrail-mcp.git guardrail-mcp

Option 2: Clone and install editable (Cross-platform: Windows, macOS, Linux)

git clone https://github.com/SECRET4422/guardrail-mcp.git
cd guardrail-mcp
pip install -e .

# Run test suite
python -m unittest discover -s tests -v

# Start MCP stdio server
guardrail-mcp

MCP client configuration

Add GuardRail to your Claude Desktop or Cursor configuration (claude_desktop_config.json):

{
  "mcpServers": {
    "guardrail": {
      "command": "guardrail-mcp",
      "args": ["serve", "--mode", "stdio"]
    }
  }
}

Or when running via direct Python executable in a local virtual environment:

{
  "mcpServers": {
    "guardrail": {
      "command": "python",
      "args": ["-m", "guardrail.server", "serve", "--mode", "stdio"],
      "cwd": "/absolute/path/to/guardrail-mcp"
    }
  }
}

CLI scan (ground truth)

python - <<'PY'
from pathlib import Path
from guardrail.hybrid_scan import hybrid_scan
r = hybrid_scan(
    Path("examples/vulnerable_sample.py").read_text(encoding="utf-8"),
    filename="examples/vulnerable_sample.py",
)
print(r["status"], r["security_verdict"], r["issue_count"], r.get("engines"))
PY

examples/ contains intentional insecure fixtures for tests and demos only (labeled in-file). Values are synthetic.

Capabilities

Area	Implementation
Secrets / high-signal patterns	rules.py, language grids
Python AST + multi-hop taint	ast_engine.py, taint.py
Tree-sitter (optional grammars)	treesitter_engine.py
Repo / git-diff scan	repo_scan.py, git_scan.py
Dependencies / OSV	deps.py (network optional)
SARIF / SBOM	sarif_export.py, sbom.py
Enterprise auth, RBAC, policy	guardrail/enterprise/
Custom rules / plugins	rule_engine.py, plugins.py

Tests

python -m unittest discover -s tests -v

Or with pytest:

pytest

Published summary: docs/test-results.md (reproduce with the command above).

Performance

Indicative micro-benchmarks are in docs/PERFORMANCE.md and benchmarks/. Re-run on your machine before relying on numbers.

Enterprise mode

Optional multi-tenant gateway (API keys/JWT, RBAC, path sandbox, audit, rate limits). See docs/ENTERPRISE.md. Do not deploy HTTP enterprise mode without authentication.

Hosted listing

Optional commercial listing: mcpize.com/mcp/guardrail.
Self-hosting the MIT core remains free. Pricing on MCPize is set in that marketplace dashboard.

Security

• Threat model: docs/SECURITY_MODEL.md
• Vulnerability reporting: SECURITY.md
• Claims policy: docs/ACCURACY.md

License

MIT — LICENSE.