Free MCP · on-device · nothing uploaded
Check PDF redactions from Claude
Add one MCP server and ask Claude "did I actually redact this?" It checks the file on your machine and tells you whether SSNs, emails, card numbers and Australian identifiers such as TFNs are still sitting under the black boxes. The PDF never leaves your computer.
Every Omitly tool for AI assistants is on the MCP page.
Quick walkthrough
Install the Omitly MCP in seconds
Download the bundle, add it to Claude Desktop, and start checking PDFs locally.
Add it to your client
Claude Desktop: one-click download, nothing else needed. Everything else: Node 20+, runs via npx — no separate install.
Before you install: check_redaction and find_sensitive_regions share a limit of 10 checks a month
and mark their results as evaluation output. locate_text and verify_redaction have no cap. Nothing is uploaded either way.
Each snippet sets OMITLY_ALLOWED_DIR: the server only
reads and writes files under this folder, so point it at wherever your PDFs live — without
it, the server defaults to whatever directory the client launched it from. (The Claude
Desktop .mcpb bundle asks you to pick this folder during install, so it needs no variable.)
The button always serves the current bundle, which may be newer than this. Pinned build: v5.5.0 .mcpb · how to check a SHA-256
SHA-256 of v5.5.0: 565d5a1815a8305b5c1c2b76e9143920ad838592d56a62fbb95a93879666f792
Or configure manually — edit claude_desktop_config.json (Settings → Developer → Edit Config), then fully quit and reopen.
{
"mcpServers": {
"omitly-leak-check": {
"command": "npx",
"args": ["-y", "omitly-mcp"],
"env": { "OMITLY_ALLOWED_DIR": "/path/to/your/pdfs" }
}
}
} One command in your terminal. The second block is optional: it installs the Omitly skill, the instructions that tell Claude when to run these checks and how to report them (masked values only, what was and was not scanned).
claude mcp add omitly-leak-check -e OMITLY_ALLOWED_DIR=/path/to/your/pdfs -- npx -y omitly-mcp
# optional: the Omitly skill (user scope)
mkdir -p ~/.claude/skills/omitly-redaction
curl -fsSL https://raw.githubusercontent.com/omitly/omitly-mcp/main/skill/SKILL.md \
-o ~/.claude/skills/omitly-redaction/SKILL.md Add to ~/.cursor/mcp.json (or .cursor/mcp.json in a project):
{
"mcpServers": {
"omitly-leak-check": {
"command": "npx",
"args": ["-y", "omitly-mcp"],
"env": { "OMITLY_ALLOWED_DIR": "/path/to/your/pdfs" }
}
}
} Add to ~/.codeium/windsurf/mcp_config.json, then reload:
{
"mcpServers": {
"omitly-leak-check": {
"command": "npx",
"args": ["-y", "omitly-mcp"],
"env": { "OMITLY_ALLOWED_DIR": "/path/to/your/pdfs" }
}
}
} Then just ask
"Did I actually redact /path/to/report.pdf?"
"Scan this PDF for any SSNs or card numbers still in the text."
"Check the exhibits in this folder and tell me which ones leak."
What a leak actually looks like
Most "redacted" PDFs are just a black rectangle drawn on top of the page — a shape, not a
removal. The original characters are still sitting in the file's content stream underneath
it, selectable and copyable the moment someone drags across the box or runs
pdftotext on the file.
How Omitly actually removes it
goes into the technical detail (content-stream removal under ISO 32000-2, not an overlay).
check_redaction on a file like that reports back, in plain text — illustrative example:
⚠️ Text-layer LEAK: this "redacted" PDF still contains 2 sensitive item(s)
(1 ssn, 1 email) in its text layer — the redaction only painted over them.
Text hidden under redaction marks (HIGH RISK):
- Page 3: ssn (***-**-4821)
- Page 3: email (j***@example.com)
Previews are masked; raw values stayed on-device. To actually remove this
data (not cover it) with independent verification + a signed audit log,
redact it in Omitly. Claude reads that back directly — which page, what kind of data, and that the box didn't actually remove it. Every preview stays masked; nothing is uploaded. You can see the same check without any setup at all in the free browser leak checker — same detector, no MCP client required.
The free tools
check_redaction Audit a 'redacted' PDF and report any PII still in the text layer underneath the boxes.
find_sensitive_regions List PII candidates (emails, SSNs, phone, card numbers) with page and coordinates.
locate_text Find exact strings — names, addresses, account refs — and their positions.
extract_pdf_text Extract a PDF's full text, page by page — PII masked by default; raw text is an explicit opt-in.
All four work with zero setup and are read-only — they never modify a file. The first two are metered: 10 checks a month between them, results marked as evaluation output; locate_text and extract_pdf_text are not. This package does not write redacted files by itself; see "Already have Omitly installed?" in the FAQ below.
This finds the problem. Omitly fixes it.
Detecting a leak is free. Removing the data for real — stripping the characters from the file, verifying nothing survives, and writing a signed audit log — means using the Omitly desktop app, where your AI can propose redactions for you to accept (see the FAQ). Detection is also pattern-based: names, addresses and image-only text aren't covered, so a clean result isn't proof of completeness. A finding means text matching a pattern is in the file; check whether it is sensitive and whether it was meant to be visible. Claude itself never does the redacting; see why asking an LLM to redact a PDF directly is unsafe.
Get OmitlyQuestions
What is the Omitly leak-check MCP?
It's a free Model Context Protocol server that gives Claude (or any MCP client) tools to inspect PDFs on your machine, out of the box: check whether a 'redacted' file still leaks PII, find sensitive regions, locate specific text, and extract a PDF's text page by page (masked by default). These run a WebAssembly build of Omitly's engine locally, so the PDF is not uploaded, and no setup beyond npx is needed. Text the tools return goes wherever your AI runs.
Does it upload my PDFs?
No. Every tool runs inside the MCP process on your computer and does not upload the PDF. A hosted AI still receives the text the tools return; use a local model if that text must not leave the machine. Once it is running, checks use no network. The npx setup contacts the npm registry when it starts, so for offline use, install the package first.
What does Claude actually see?
The file stays on your computer — the results don't. Everything a tool returns — masked previews, findings, and the page-by-page text extract_pdf_text produces (raw text only if you explicitly pass masked: false) — goes back to your AI client like any other tool result, and from there it's covered by that client's privacy terms, not Omitly's. Masking is pattern-based, too: recognised identifiers (emails, SSNs, phone and card numbers) are masked, but a name, an address or a sensitive passage that matches no pattern comes back as plain text.
Can it redact a PDF for me?
No. The tools above are read-only: they tell you a redaction leaked, or extract text, but they never modify a file. To remove the data, with verification and a signed audit log, use the Omitly desktop app. With the app open, your AI can propose redactions and you accept them in the app; see Connect your AI in the help centre.
Already have Omitly installed? What else can this MCP do?
With the app open and Accept proposals in Omitly switched on, your AI can propose redactions and you accept them in the app. This MCP cannot write a redacted file by itself: the native engine it would need is not distributed, so redact_pdf, redact_by_entity and create_pdf return a 'not configured' message. It cannot verify seals either; use Verify in the desktop app. verify_redaction is free and re-scans a file for you.
Do I need to install anything?
For Claude Desktop, no — download the .mcpb bundle above and drag it in. For any other client (Claude Code, Cursor, Windsurf), just Node 20+ — the server runs via npx, so there's no separate install step.
How do I check the .mcpb I downloaded?
The download button serves the current bundle. The pinned link beside it points at one specific version, and that version's SHA-256 is printed underneath it — run `shasum -a 256` (macOS) or `Get-FileHash -Algorithm SHA256` (Windows) on the file you downloaded and compare the two strings. The same version, URL and SHA-256 are published as JSON at https://releases.omitly.app/mcp/latest.json. Note the printed hash describes the pinned version, not the always-current button: if a newer bundle has published since, the button's file will hash differently, which is expected.
Is it really free?
Yes — free, no licence, nothing uploaded. Two of the tools are metered: check_redaction and find_sensitive_regions share a limit of 10 checks a month, and their results carry an evaluation mark. locate_text, extract_pdf_text and verify_redaction have no cap. The count is kept in a local file and never reported anywhere. Redacting a file is done in the Omitly desktop app, on its own trial and licence terms.