Scans Agent2Agent protocol implementations for webhook SSRF, missing authentication, and credential exposure in TypeScript and Python code.
Also known as GuardBee/guardbee-mcp, @guardbee/mcp-a2a-auditor
- Stars
- 2
- Tools
- 4
- Downloads, 30 days
- 1,814
- Latest version
- 0.1.1
- Last commit
- Nov 12, 2026
- License
- MIT
What is a2a-auditor MCP?
a2a-auditor scans Agent2Agent protocol implementations for webhook SSRF, missing authentication, and credential exposure in TypeScript and Python code. It exposes 4 tools, such as scan_text, scan_file, scan_directory.
Summary generated from the project's README
About a2a-auditor MCP server
Scanning Agent2Agent protocol implementations for security vulnerabilities is what a2a-auditor enables for developers working with TypeScript and Python. This MCP server identifies specific anti-patterns found in reference SDKs, such as webhook SSRF, missing authentication, and credential exposure. It is designed to be used locally via an MCP client like Claude Desktop or integrated into CI/CD pipelines through a command line interface.
Users can perform audits using tools like scan_text, scan_file, or scan_directory to recursively check entire codebases. The server detects critical issues like the absence of security schemes in an Agent Card or the direct fetching of webhook URLs without host validation. It also flags high-severity risks where API keys are left inside publicly accessible metadata. The system supports SARIF 2.1.0 output for integration with GitHub Code Scanning and allows configuration through a guardbee.yml file to set failure thresholds and file exclusions.
What you can do with it
- 1Audit local agent source codeA developer can scan their entire project directory to find security flaws in their A2A implementation.
Written from the project's README. Every use case is checked against a quote from it.
How to install a2a-auditor MCP
Formats verified 10 Nov 2026Add to CursorAdd to GooseInstall a2a-auditor MCP in Claude Code
$ claude mcp add a2a-auditor -- npx -y @guardbee/mcp-a2a-auditorInstall a2a-auditor MCP in Cursor
~/.cursor/mcp.json
{
"mcpServers": {
"a2a-auditor": {
"command": "npx",
"args": [
"-y",
"@guardbee/mcp-a2a-auditor"
]
}
}
}Install a2a-auditor MCP in VS Code
.vscode/mcp.json
{
"servers": {
"a2a-auditor": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"@guardbee/mcp-a2a-auditor"
]
}
}
}Install a2a-auditor MCP in VS Code CLI
$ code --add-mcp '{"name":"a2a-auditor","type":"stdio","command":"npx","args":["-y","@guardbee/mcp-a2a-auditor"]}'Install a2a-auditor MCP in Claude Desktop
claude_desktop_config.json
{
"mcpServers": {
"a2a-auditor": {
"command": "npx",
"args": [
"-y",
"@guardbee/mcp-a2a-auditor"
]
}
}
}Install a2a-auditor MCP in Codex
$ codex mcp add a2a-auditor -- npx -y @guardbee/mcp-a2a-auditorInstall a2a-auditor MCP in Codex config.toml
~/.codex/config.toml
[mcp_servers.a2a-auditor] command = "npx" args = ["-y", "@guardbee/mcp-a2a-auditor"]
Install a2a-auditor MCP in Gemini CLI
$ gemini mcp add -s user a2a-auditor npx -- -y @guardbee/mcp-a2a-auditorInstall a2a-auditor MCP in Gemini settings.json
~/.gemini/settings.json
{
"mcpServers": {
"a2a-auditor": {
"command": "npx",
"args": [
"-y",
"@guardbee/mcp-a2a-auditor"
]
}
}
}a2a-auditor MCP tools 4
Detected from the READMEscan_textscan_filescan_directorylist_patterns
| scan_text | β | |
| scan_file | β | |
| scan_directory | β | |
| list_patterns | β |
Automated checks
Ran Nov 13, 2026Full report βPassed our automated checks. That is not a security audit: read the code and give the server the narrowest access it needs. How checks work
- Source repository linked and reachablePass
- Open source license (OSI)Pass
- Commit or release in the last yearPass
- Package name not close to a popular packagePass
- Remote endpoint uses HTTPSNot checked
- Destructive tools are labelledPass
- No hidden instructions to the modelPass
- No critical advisory in the package or its dependenciesPass
README
View on GitHub@guardbee/mcp-a2a-auditor
π¬π§ English | πΉπ· TΓΌrkΓ§e | π¨π³ δΈζ | π―π΅ ζ₯ζ¬θͺ | π°π· νκ΅μ΄ | π§π· PortuguΓͺs | πͺπΈ EspaΓ±ol | π©πͺ Deutsch
An MCP (Model Context Protocol) server that scans an agent's own Agent2Agent (A2A) protocol implementation for anti-patterns β 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 β silently meaning "no auth" if you don't pass them.
Telemetry is off by default. With
GUARDBEE_TELEMETRY=1it sends the server, tool, success, duration and short parameter values such as a file path; longer strings and fields likecontent,tokenorpasswordare replaced with a placeholder. See@guardbee/mcp-telemetry.
Claude βββΊ a2a-auditor βββΊ Your A2A agent's source
β
ββ webhook-ssrf (push-notification webhook URL fetched with no allowlist)
ββ missing-authentication (Agent Card / request handler requires no auth at all)
ββ 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, ...) β 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 β 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 β 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 β 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
urlvariable back to apushConfig.url-shaped assignment before flagging thefetch()that uses it, and the credential check extracts the full Agent Card object span (brace-matched) before searching inside it for a literal secret β so it doesn't just flag anyapiKey:anywhere in the file - SARIF 2.1.0 output β CI/CD integration (GitHub Code Scanning, etc.)
- Config file support via
guardbee.yml - 18 unit tests β a positive and a negative (false-positive) scenario for every pattern, including a regression test for a real bug found during development (
\bAgentCarddoesn't match insidemovieAgentCardβ no word boundary between a lowercase and uppercase letter β 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 β 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 β 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 β a
urlvariable 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 multipleObject.assign()-style merges instead of one literal. - Naming-convention dependent. The checks look for identifiers shaped like
pushConfig.url,*AgentCard,UserBuilder.noAuthenticationβ 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 β 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 β on the repository helps others find it.
a2a-auditor MCP questions
How do I add a2a-auditor MCP to Claude Code?
Run claude mcp add a2a-auditor -- npx -y @guardbee/mcp-a2a-auditor in your terminal. Add --scope user to make it available in every project.
How do I add a2a-auditor MCP to Cursor?
Open ~/.cursor/mcp.json and add the server under mcpServers, using the Cursor config shown in the install section above.
What tools does a2a-auditor MCP provide?
It exposes 4 tools, including scan_text, scan_file, scan_directory, list_patterns.
Does it support both TypeScript and Python?
Yes, the server checks both TypeScript and Python A2A code to account for different SDK idioms.
Can I integrate this into my CI/CD pipeline?
Yes, it supports SARIF 2.1.0 output, which allows for integration with tools like GitHub Code Scanning.
Is telemetry enabled by default?
No, telemetry is turned off by default.