Most "ATS scanners" online give you a 0-100 score that's invented. After talking to recruiters at three companies on different ATS platforms, I confirmed: there is no real ATS score. Different parsers extract different fields, and the scoring tools just guess.

So I built a tool that does it differently — and shipped it as an MCP server so any agent (Claude Code, Cursor, Windsurf, Zed, etc.) can use it inline.

Repo: goofypluto999/cv-mirror-mcp
License: MIT.

What it actually does

Three tools exposed over the Model Context Protocol:

• analyze_cv(path) — full report. Takes a CV (PDF or DOCX), returns per-vendor lint findings across Workday, Greenhouse, Lever, Taleo, iCIMS with severity (error / warn / info) and concrete fixes.
• lint_for_vendor(path, vendor) — single-vendor lint. Faster when the user only cares about, say, Workday.
• get_express_url() — returns the URL of a visual web companion (live in-browser parse view) for users who want the side-by-side overlay.

Sample output:

[ERROR] Workday
  - WORKDAY_MULTI_COLUMN: 35% of lines look multi-column. Workday's parser
    reads left-to-right and interleaves both columns into one stream.
    Fix: Convert to single-column layout. Move sidebars (Skills, Tools)
    above or below the main content.

[OK] Greenhouse
  No issues detected by the simulated parser.

[WARN] Lever
  - LEVER_HEADER_FOOTER: Header/footer-like text detected ("Page 1 of 2").
    Lever historically drops content placed in PDF headers/footers.

Why per-vendor instead of a single score

The five top enterprise ATSes do parsing differently:

• Workday reads PDFs in document-stream order — multi-column layouts get interleaved
• Greenhouse strips most emoji codepoints; "Projects 🚀" can lose its surrounding context
• Lever historically drops content in PDF headers/footers
• Taleo prefers Month-Year date format; ISO dates often fail to populate employment durations
• iCIMS has the worst multi-column handling — sidebar Skills sections often merge with the line above

A single score collapses all of that into noise. Per-vendor surfaces what each one actually sees.

Stack

• pdf-parse for PDF text extraction
• mammoth for DOCX
• @modelcontextprotocol/sdk for the MCP server transport (stdio)
• Pure-JS vendor simulators in src/lint.mjs — no AI inference, just heuristics derived from public vendor docs
• 19 unit tests covering signal extraction and per-vendor rule firing

The vendor rules are intentionally conservative. Every rule cites its public source in docs/vendor-sources.md. I'd rather miss an issue than false-positive on a real job page.

Install + use

npm install -g cv-mirror-mcp

Add to your MCP client config:

{
  "mcpServers": {
    "cv-mirror": {
      "command": "npx",
      "args": ["-y", "cv-mirror-mcp"]
    }
  }
}

Restart the client. Then ask your agent:

Scan my CV at ~/Documents/resume.pdf and tell me what each ATS would do to it.

The agent calls analyze_cv, the lint engine runs locally inside the agent's process — no upload, no telemetry — and the agent surfaces the per-vendor breakdown.

Privacy by architecture

The express path (web tool) runs entirely client-side via pdf.js. The MCP server runs locally inside your agent's process. There is no upload endpoint and no network call from the lint engine. You can verify in the source — it's MIT licensed.

This isn't a marketing claim. The architecture makes the data exfiltration impossible, not just unlikely.

Roadmap

• [ ] BambooHR simulator
• [ ] SmartRecruiters simulator
• [ ] Real-time JD scraping for major ATS public job boards (so the user can also pass a JD URL and get keyword-match findings)
• [ ] Browser extension that overlays parser output on any job application form

Sister project (paid, optional)

The same team also makes Vantage AI — once your CV is past the parser, the next problem is the application: tailored cover letter, mock interview, fit analysis. That's a paid SaaS, separate from this MCP server. Genuinely free signup with 3 free analyses included if you want to try it.

The MCP server stands alone — it's MIT, it's free, it's not a teaser.

Contributing

PRs welcome. Particularly interested in:

• More vendor simulators (BambooHR, SmartRecruiters, JazzHR, Recruitee, etc.)
• Updated rules when vendors change their parsing behaviour (open an issue with the source link)
• Translations of the lint output

The repo is at goofypluto999/cv-mirror-mcp. Built solo over a couple of weeks.