Automation Auditor MCP

One MCP server, several real automation export formats. Built and maintained by Ahmad Yar.

This is an early-stage implementation. See the PyPI package and MCP Registry entry for current publishing and version details. Host compatibility and real-client MCP behavior should be verified before production use. It performs read-only static checks on supplied text; it never reads your repo on its own, contacts automation accounts, runs JS, or connects to API endpoints.

Supported inputs and honest limitations

"Any automation platform" is not possible without an adapter for its native format. Unsupported platforms are reported as unsupported, rather than passing a meaningless check. You can add adapters to the registry in core.py and corresponding tests. Make, Zapier and Pipedream are not advertised as supported in v0.1.

Related MCP server: toolfence

Install

Install with Python 3.11+:

pip install automation-auditor-mcp

Configure a compatible MCP client:

{
  "mcpServers": {
    "automation-auditor": {
      "command": "uvx",
      "args": ["automation-auditor-mcp"]
    }
  }
}

uvx requires uv. Confirm the package's current published version on PyPI and test the configuration with your MCP host before production use.

For development from source:

python -m pip install -e .
python -m unittest discover -s tests -v
automation-audit examples/n8n-valid.json --platform n8n
automation-auditor-mcp

The last command runs the MCP stdio server and waits for a connected client. Use a host or MCP Inspector to communicate with it; starting from an ordinary terminal should not print output.

MCP tool API

• audit_workflow(platform, source): Pass the full text contents of an export or Trigger.dev source file. platform is auto, n8n, dify, trigger, trigger.dev, or activepieces. Max 250,000 characters. The server does not read files by path.

• audit_openapi(spec_json): JSON OpenAPI document; uses the existing ahmad-yar-gpt-tools library.

• preflight_url(url): Pure URL screening; never performs DNS lookup or HTTP request. A permitted syntax becomes manual_review because real fetching requires request-time controls.

• list_supported_platforms(): Exact supported formats and known limitations.

Findings include a stable rule ID, severity, location, brief message, recommendation and confidence. They intentionally do not echo node names, raw URLs, credentials or source snippets. A report of no_findings_detected is not proof of safety.

Example: detect a broken n8n connection

The MCP client passes exported workflow JSON to audit_workflow("n8n", source). If a connection points to a deleted node, the structured report includes rule N8N009 with severity error. The assistant can explain the error and request the user to fix it. This tool never edits workflow files itself.

Tests

python -m unittest discover -s tests -v

Test fixtures use only fictional metadata; no API keys or customer information. The core format inspectors require no network and do not import MCP at test time; MCP SDK integration requires installing dependencies.

Release and registry maintenance checklist

1. Keep the public source repository and installation instructions synchronized with released artifacts.

2. Run all tests and verify MCP Inspector calls in a real MCP client for each release. Improve Trigger.dev analysis with a TypeScript AST parser before describing it as comprehensive.

3. For each release, verify the published version and metadata on PyPI, including its source and homepage links.

4. Preserve the README marker mcp-name: io.github.ahamdjin/automation-auditor-mcp inside the PyPI long description for Registry package ownership verification.

5. When releasing a new version, validate server.json using the official MCP Registry publisher, and keep its PyPI identifier, version and stdio transport accurate.

6. Verify Registry metadata after each release using the official Registry API. The Registry hosts metadata, not the package itself.

The Registry's official publishing requirements explain PyPI ownership verification.

Security, licensing and attribution

No source content is intentionally stored, sent externally, executed, or used for model inference. The MCP host may log inputs or reports separately. YAML aliases are refused, duplicate JSON/YAML fields are rejected, and input size is bounded. Trigger.dev regex checks have low confidence; a source audit cannot guarantee runtime behavior. Never treat URL preflight as SSRF prevention.

License: Apache-2.0. Upstream ahmad-yar-gpt-tools also uses Apache-2.0.

Source (after repository creation): https://github.com/ahamdjin/automation-auditor-mcp Creator: https://www.ahmadyar.co/