InRedactr
Your documents never leave this machine.

Strip the names out before the AI ever sees them

Paste a document, get back a version that is safe to hand to ChatGPT, Claude or anyone else. InRedactr finds names, addresses, Medicare and tax file numbers, cards, credentials and more, and replaces them. Then paste the AI’s answer back and your real names return. No cloud, no account, no telemetry.

Windows 10 / 11 x64  ·  One installer  ·  Nothing else to install

0
Documents uploaded
Network access is switched off inside the app. You can test it yourself.
116
Detectors
Checksum-verified, not just pattern-matched.
1
Installer
Python, the language model, everything. Nothing else to fetch.
Free
No subscription
No account, no licence key, no nag screen.

Why it exists

Pasting into an AI is a disclosure

The moment a client name, a patient record or a staff spreadsheet goes into a chat box, it has left your control. Most people know this and paste anyway, because manually scrubbing a document is tedious and easy to get wrong. InRedactr makes the safe version the quick one.

What you have

From: Sarah Whitfield <s.whitfield@meridian-health.com.au>

Please review the file for James Okonkwo,
DOB 14/03/1978, Medicare 2123456701.
Next of kin is his sister Priya
on 0412 345 678.

What you paste

From: [PERSON_1] <[EMAIL_1]>

Please review the file for [PERSON_3],
DOB [DOB_1], Medicare [MEDICARE_1].
Next of kin is his sister [PERSON_4]
on [PHONE_1].

The AI can still follow that PERSON_1 wrote to PERSON_3 about PERSON_4. It just cannot tell you who any of them are.

How it works

Three steps, and the third is the clever one

Most redaction tools stop at step two and leave you with an answer full of square brackets.

STEP 01

Paste

Paste text, drop a file, or press Ctrl+Alt+R to redact whatever is on your clipboard without even opening the window. Word, PDF, Excel, email and screenshots all work.

STEP 02

Review

Everything found is colour-coded and listed. Hover to see why it was detected, click to reject a false positive, or change how any single item is replaced. Then copy.

STEP 03

Restore

Paste the AI's reply back into InRedactr and the real names return. The mapping is encrypted, stays on your machine, and expires on a timer you set.

See it working

Captured from the application itself

On a document invented for the purpose. Every identifier below is synthetic but checksum-valid, so the app treats it exactly as it would treat a real one.

The InRedactr Redact tab: source document on the left with colour-coded highlights, redacted output on the right, and a findings table below.

Source and result, side by side Ten colour-coded categories. Look at the findings table: "Sarah Whitfield" appears nine times under three different spellings and is still one person, one token. The risk meter reads CRITICAL because a Medicare number and an AWS key are both in the document.

What it finds

Five layers, because one is never enough

A regular expression alone cannot tell a tax file number from an invoice number, and a wordlist has never heard of your colleague's surname. InRedactr runs five passes and reconciles them.

Verified identifiers

Tax File Number, ABN, ACN, Medicare, NHS, National Insurance, SSN, IBAN, Canadian SIN, Aadhaar, Indian PAN, NZ IRD, Singapore NRIC, payment cards, IMEI, VIN and Bitcoin. Each one is checksum-verified, so a near-miss is ignored.

Context, not just shape

Nine digits are nothing until something nearby calls them a Tax File Number. A date beside "DOB" is treated very differently from a date in a meeting note.

Names it has never seen

A bundled language model finds people, organisations and places that no wordlist contains. It runs on your machine, and it ships in the installer.

Structure and columns

JSON, XML, YAML and CSV are read by field name, so a value under a key called "ssn" is caught even if it looks like nothing in particular.

Hidden inside encodings

Base64, hex, URL-encoding and gzip are decoded and scanned. A token payload with an email address buried in it does not get past.

Lookalike characters

Cyrillic letters that look Latin, zero-width characters and spaced-out digits are all folded before anything is matched.

Credentials and keys

AWS, Azure, GCP, GitHub, Slack, Stripe and OpenAI keys, JSON Web Tokens, private keys, database connection strings and plain passwords.

Quasi-identifiers

The one nobody thinks of. "The CFO, based in Newcastle, born 1978" has no name and no number, and identifies exactly one person. InRedactr says so.

Your own words

Client names, project codenames and internal system names that no general tool could know about. Add them once and they are always removed, with a matching list for the words it keeps flagging that identify nobody.

Person Organisation Location Email Phone Government ID Financial Technical Date Your own terms

Privacy

A privacy tool that phones home is not a privacy tool

So InRedactr does not have the ability. This is enforced in the code, not stated in a policy.

Nothing you redact is transmitted

Outbound connections, DNS lookups and HTTP requests are all switched off inside the process when the app starts. No document text, no file names and no usage data can leave, because the ability to send them is removed.

One exception, and you control it

If you leave update checking on, InRedactr asks GitHub once a day whether a newer version exists. That request carries nothing about you, your machine or your documents, nothing installs by itself, and switching it off in Settings means nothing is ever contacted.

Test it yourself

One button in Settings attempts four real connections and shows you each being refused, with the error. You do not have to take anyone's word for it.

Encrypted for your account

The mapping that lets you restore an answer is encrypted with Windows DPAPI under your user account, expires on a timer, and is wiped when you close the app. Copy the file to another machine and it decrypts to nothing.

One-way mode

If you never need to restore anything, switch it on and no mapping is ever created. Nothing is stored anywhere.

An audit log without the data

It records that a redaction happened, of what kinds, and how many. It never records the values. Rows are hash-chained so an edited entry is detectable.

Real PDF redaction

The text is removed from the file, not covered with a black rectangle. Try selecting and copying from a redacted page: there is nothing there.

Your client list stays yours

The terms you add to always be removed are themselves encrypted on disk. A list naming your clients is a client list, so it is never stored in the clear and never included in a rule set you share with someone else.

Nothing left in Temp

Working files go to your own application data folder, never the system temporary directory, and the mapping file is overwritten before it is deleted rather than simply unlinked.

The part most pages leave out

It will not catch everything

No de-identification tool does, and any tool that tells you otherwise is selling you something. InRedactr is built to fail in the safer direction: it over-detects, shows you exactly what it found and why, and makes rejecting a false positive a single click.

It also re-scans its own output from scratch after every redaction and refuses to let you copy the result if an original value somehow survived. Read the output before you share it. The tool assists that judgement; it does not replace it.

Download

Get InRedactr

One installer. Windows 10 or 11, 64-bit. Everything is included, including the language model, so there is nothing else to fetch and nothing to configure.

Free  ·  No account  ·  No subscription  ·  No telemetry

InRedactr © 2026 Fady Azzi 116 detectors · 9 strategies · 12 profiles