# omitly-leak-check-mcp

A free, local **MCP server** that answers one question: *did your PDF redaction
actually work, or is the data still sitting in the file underneath the black
boxes?*

Most tools "redact" by drawing a rectangle over text — the characters stay in the
PDF and are trivially recoverable. This server re-extracts the text layer and
flags any emails, SSNs, phone or card numbers that survived.

- **Zero install** — `npx omitly-leak-check-mcp`. No native binary, no qpdf.
- **Nothing uploaded** — the engine is a WebAssembly build of Omitly's detector
  that runs inside the MCP process. Your PDF never leaves the machine.
- **Masked** — results show `•••-••-6789`, never the raw value.

## Use with Claude

Add to your MCP client config (e.g. Claude Desktop `claude_desktop_config.json`):

```json
{
  "mcpServers": {
    "omitly-leak-check": {
      "command": "npx",
      "args": ["-y", "omitly-leak-check-mcp"]
    }
  }
}
```

Then ask: *"Did I actually redact /path/to/file.pdf?"*

## Tools

| Tool | What it does |
|------|--------------|
| `check_redaction` | Audit a "redacted" PDF; report surviving PII (masked). |
| `find_sensitive_regions` | List PII candidates with page + coordinates. |
| `locate_text` | Find exact strings (names/addresses) and their positions. |

All three are **detect-only and read-only.** They never modify a file.

## This finds the problem. Omitly fixes it.

Detecting a leak is free. *Removing* the data for real — with independent
verification that nothing survives, plus a signed audit log — is the
[Omitly](https://omitly.app) desktop app. Detection is also pattern-based: names,
addresses, and text inside scanned images aren't covered, so a clean result is
not a guarantee of completeness.

## Develop

```bash
npm install && npm run build
npm run smoke -- /path/to/test.pdf   # prove the wasm loads + scans
```

The wasm in `wasm/` is built from `crates/leakcheck-wasm` in the Omitly repo
(`wasm-pack build --target web`).