@guardbee/mcp-a2a-auditor

@guardbee/mcp-a2a-auditor

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

An MCP (Model Context Protocol) server that scans an agent's own Agent2Agent (A2A) protocol implementation for anti-patterns — not invented from the spec text, but read directly out of the reference SDKs (@a2a-js/sdk, a2a-sdk) and their own official sample code.

A2A doesn't have an MCP-style "Security Considerations" section yet to quote. So instead of guessing, this package's patterns come from three things found in the reference SDK source itself: the official push-notification sender fetches a client-supplied webhook URL with no host check at all, the official sample agent ships with an empty security scheme and a literal noAuthentication request handler, and the Python SDK's AgentCard(...) constructor treats its two security kwargs as fully optional — silently meaning "no auth" if you don't pass them.

This package sends usage telemetry by default (tool name + short parameters, scanned code is never included — see @guardbee/mcp-telemetry). Disable with GUARDBEE_TELEMETRY=0.

Claude ──► a2a-auditor ──► Your A2A agent's source
              │
              ├─ webhook-ssrf            (push-notification webhook URL fetched with no allowlist)
              ├─ missing-authentication  (Agent Card / request handler requires no auth at all)
              └─ credential-exposure     (literal credential inside publicly-served Agent Card metadata)

Why these three specifically

Webhook SSRF because the reference DefaultPushNotificationSender does exactly const url = pushConfig.url; ... fetch(url, ...) — pushConfig.url is set by whichever client subscribed to task updates, and the SDK fetches it with zero host validation. A caller can point a server running this code at an internal service, a cloud metadata endpoint, or anything else reachable from the server's network.

Missing authentication because the SDK's own sample agent (src/samples/agents/movie-agent) ships with securitySchemes: {}, securityRequirements: [], and userBuilder: UserBuilder.noAuthentication — a copy-paste-friendly shape that looks like a placeholder but is, functionally, "accept every request from anyone." The Python SDK has the same failure mode via absence rather than an explicit empty value: AgentCard(...) never requires security_schemes/security_requirements, so a real agent built by following the reference examples can end up with no authentication and no error telling you so.

Credential exposure because an Agent Card is served publicly — anyone who can reach /.well-known/agent-card.json (or the equivalent discovery response) can read it, no authentication required to fetch the card itself. A literal API key sitting in that JSON is effectively public.

Features

• 3 categories: webhook-ssrf, missing-authentication, credential-exposure
• Both TypeScript and Python A2A code — the SDKs use different idioms for the same failure (empty object vs. absent kwargs), so this checks both, not just one language
• Two of the six checks track data flow across a few lines instead of matching a single call site: the webhook check follows a url variable back to a pushConfig.url-shaped assignment before flagging the fetch() that uses it, and the credential check extracts the full Agent Card object span (brace-matched) before searching inside it for a literal secret — so it doesn't just flag any apiKey: anywhere in the file
• SARIF 2.1.0 output — CI/CD integration (GitHub Code Scanning, etc.)
• Config file support via guardbee.yml
• 18 unit tests — a positive and a negative (false-positive) scenario for every pattern, including a regression test for a real bug found during development (\bAgentCard doesn't match inside movieAgentCard — no word boundary between a lowercase and uppercase letter — fixed to match the class name as a suffix instead). End-to-end verified against fixtures with all patterns deliberately planted in both TypeScript and Python (caught all of them), and against the entire guardbee-mcp monorepo (461 files — zero findings in any real code; the only matches were this package's own test fixtures, pattern-definition source, and manifest description text)

Quick Start

Claude Desktop / MCP Client

{
  "mcpServers": {
    "guardbee-a2a-auditor": {
      "command": "npx",
      "args": ["-y", "@guardbee/mcp-a2a-auditor"]
    }
  }
}

CLI (CI/CD)

npx @guardbee/mcp-a2a-auditor scan ./src --fail-on=high --format=sarif > results.sarif

MCP Tools

Tool	Description
scan_text	Scans a code snippet
scan_file	Scans a single file
scan_directory	Recursively scans a directory
list_patterns	Lists all supported patterns by category

Detected Patterns

Category	Pattern	Severity
webhook-ssrf	webhook_url_direct_fetch_no_allowlist	critical
webhook-ssrf	webhook_url_indirect_fetch_no_allowlist	critical
missing-authentication	no_authentication_user_builder	critical
missing-authentication	empty_agent_card_security	high
missing-authentication	python_agent_card_no_auth	high
credential-exposure	agent_card_credential_in_metadata	high

Configuration (guardbee.yml)

a2a-auditor:
  fail-on: high       # any | critical | high | medium | none
  max-files: 5000
  exclude:
    - ".test.ts"

Limitations (by design)

• Heuristic, not a full A2A conformance test. This finds specific anti-patterns taken from the reference SDK's own source, not a live protocol conformance check against a running agent. A2A is a young, fast-moving spec — new failure modes will show up over time that aren't covered here yet.
• The data-flow checks (indirect webhook fetch, Agent Card credential span) use a bounded text window and bracket-matching, not a real parser — a url variable assigned through a more indirect chain (e.g. passed through a helper function) won't be traced correctly, and neither will an Agent Card object spread across multiple Object.assign()-style merges instead of one literal.
• Naming-convention dependent. The checks look for identifiers shaped like pushConfig.url, *AgentCard, UserBuilder.noAuthentication — code that renames these away from the SDK's own conventions won't be recognized, the same limitation every pattern-based scanner in this family has.

Development

npm run build
npm test             # 18 unit tests

License

MIT — GuardBee