@guardbee/mcp-ai-code-scanner

@guardbee/mcp-ai-code-scanner

🇬🇧 English | 🇹🇷 Türkçe | 🇨🇳 中文 | 🇯🇵 日本語 | 🇰🇷 한국어 | 🇧🇷 Português | 🇪🇸 Español | 🇩🇪 Deutsch

An MCP (Model Context Protocol) server that scans your codebase for security patterns commonly found in LLM/AI integrations.

While secret-scanner looks for leaked secrets/API keys, this package looks for patterns: how much a model's output is trusted, how privileged the tools it can call are, what data is sent to a third-party LLM — risks that a static credential scan can't catch.

Telemetry is off by default. With GUARDBEE_TELEMETRY=1 it sends the server, tool, success, duration and short parameter values such as a file path; longer strings and fields like content, token or password are replaced with a placeholder. See @guardbee/mcp-telemetry.

Claude ──► ai-code-scanner ──► Your codebase
              │
              ├─ Client-exposure    (dangerouslyAllowBrowser: true)
              ├─ Output handling    (eval(llmOutput), unvalidated JSON.parse)
              ├─ Excessive agency   (an agent tool named execute_command)
              ├─ Data privacy       (email/national ID interpolated directly into a prompt)
              └─ Prompt injection   (raw request data mixed into a system prompt)

Features

• 10 patterns, 5 categories — client-exposure, output-handling, excessive-agency, data-privacy, prompt-injection
• Every finding includes why it's risky and what to do about it (recommendation) — not just "found it"
• SARIF 2.1.0 output — CI/CD integration (GitHub Code Scanning, etc.)
• Config file support via guardbee.yml
• Custom rule packs — extend detection with org-specific patterns (KVKK/GDPR field names, internal tool names) via JSON files, no fork required
• 36 unit tests — a positive and a negative (false-positive) scenario for every built-in pattern, plus custom-rule loading/validation

Quick Start

Claude Desktop / MCP Client

{
  "mcpServers": {
    "guardbee-ai-code-scanner": {
      "command": "npx",
      "args": ["-y", "@guardbee/mcp-ai-code-scanner"]
    }
  }
}

CLI (CI/CD)

npx @guardbee/mcp-ai-code-scanner scan ./src --fail-on=high --format=sarif > results.sarif

MCP Tools

Tool	Description
scan_text	Scans a given text/code snippet
scan_file	Scans a single file
scan_directory	Recursively scans a directory (node_modules, .git, dist skipped automatically)
list_patterns	Lists all supported patterns by category

Detected Patterns

Category	Pattern	Severity	What it means
client-exposure	openai_dangerously_allow_browser	critical	OpenAI key leaked to the browser
client-exposure	client_bundled_ai_api_key	critical	An AI key enters the client bundle via NEXT_PUBLIC_/VITE_/REACT_APP_
output-handling	eval_llm_output	critical	Model output executed as code via eval()/Function()
output-handling	exec_llm_output	critical	Model output passed to a shell command (command injection)
output-handling	llm_output_dangerously_set_inner_html	high	Model output rendered as raw HTML (XSS)
output-handling	llm_json_no_validation	medium	Model output JSON.parse'd without schema validation
excessive-agency	excessive_agency_tool_name	high	A tool that grants an agent shell/code execution
excessive-agency	unbounded_agent_loop	medium	An agent loop with no iteration bound (cost/DoS)
data-privacy	pii_field_in_llm_prompt	high	Data like email/national ID/card number goes directly into a prompt
prompt-injection	unsanitized_input_in_system_prompt	medium	Raw request/user data mixed into a system prompt

These are heuristic findings — a static text-pattern scan, not a full AST analysis. Designed for a low false-positive rate, but every finding should still be reviewed manually.

Configuration (guardbee.yml)

ai-code-scanner:
  fail-on: high       # any | critical | high | medium | low | none
  max-files: 5000
  rules-dir: .guardbee/rules   # custom rule packs, see below
  exclude:
    - ".test.ts"
    - "fixtures/"

Custom rules

Built-in patterns cover common AI/LLM integration mistakes, but field names and internal conventions (e.g. which prop names carry KVKK-regulated data) are project-specific. Drop one or more JSON files into .guardbee/rules/ (configurable via rules-dir) and they're merged with the built-ins at scan time — for both the CLI and the MCP server (list_patterns marks them (custom)).

A file holds a single rule object or an array of them:

{
  "id": "kvkk_musteri_tc_no",
  "name": "musteriTcNo interpolated into a prompt",
  "category": "data-privacy",
  "pattern": "musteriTcNo",
  "flags": "gi",
  "severity": "high",
  "recommendation": "Mask the national ID before it enters the prompt."
}

category is one of client-exposure | output-handling | excessive-agency | data-privacy | prompt-injection | cost-control; severity is one of critical | high | medium | low. pattern is a JS regex source string (flags defaults to "g"). A rule id that collides with a built-in or another custom rule is rejected — the scanner keeps running but reports the problem on stderr and skips that entry.

Development

npm run build
npm test             # 36 unit tests

License

MIT — GuardBee

Feedback

Questions, ideas, or a false positive? Start a thread in GitHub Discussions or open an issue. If GuardBee is useful to you, a ⭐ on the repository helps others find it.