The AI chat for teams that can't share everything.

SOWA Privacy Chat is a desktop chat application for macOS and Windows, with detection built into the send path. Prompts are scrubbed on your machine, then sent to ChatGPT, Claude, Gemini or Grok with your own API key.
Included at no extra cost in the Business and Enterprise plans.

SOWA Privacy Chat app icon

How a prompt travels through the app

Detection in the composer

The SOWA Privacy Chat composer. A bar above the prompt reports four detected items, broken down into address, blacklisted term, email and person. In the prompt itself, the name Maria Schmidt, the address Hauptstraße 42 in Berlin, the reference BPA-2026-08834 and the address maria.schmidt@web.de are highlighted in colour.
The app marks the personal data it finds while you are still typing.

Review before anything leaves the machine

The “Review sensitive items” dialog. On the left the original text, on the right the same text carrying the placeholders PERSON_1, ADDRESS_2, BLACKLIST_3 and EMAIL_1@masked.local. Below, the four items are listed individually, each with a switch to redact or pass through.
The original on the left, the text that will be sent on the right.

There and back

What reaches the provider Chat thread showing placeholders: the model’s reply refers to PERSON_1, ADDRESS_2, BLACKLIST_3 and EMAIL_1@masked.local.
What you see The same chat thread with original values restored: Maria Schmidt, Hauptstraße 42 in 10115 Berlin, BPA-2026-08834 and maria.schmidt@web.de.

One conversation, two views. The model answers placeholders; one click on “Originals” shows your original values.

Captured from SOWA Privacy Chat V1.0.0.0. The data shown is fictitious.

What it does

Every item below runs in-process on your machine.

Local PII pipeline

Regex, a named-entity model and an optional local LLM detector run in sequence. Eight ready-made profiles for banking, medical, insurance, HR, legal and IT ship with it, and you can add rules of your own.

Anonymisation dictionary

Placeholders round-trip: the reply comes back and your original values are mapped in again, so redaction never breaks the thread. Scope per thread or global.

Privacy Owl

A floating widget showing live detection status, with a side-by-side review dialog before anything is sent.

Attachment scrubbing

PDF, DOCX, XLSX and text formats are scanned and redacted before they reach a provider. DOCX de-anonymisation preserves formatting.

Audit log

A SHA-256-chained log with integrity check, ten event types and JSON or CSV export. Original values optionally hashed.

Your own API keys

ChatGPT, Claude, Gemini and Grok. Base URL, model and key are configurable per provider; keys are stored locally.

File uploads too

PDF DOCX XLSX + text formats

Works with

ChatGPT Claude Gemini Grok
  • No guarantee of completeness. No detection model reaches 100% recall. The pipeline is tuned for precision; values it misses are possible, and we make no promise of “zero leaks”.
  • False positives happen. Sometimes a harmless value gets masked. The review dialog lets you correct it by hand and an exception list makes it stick; on Business and Enterprise, administrators can tune this centrally.

Get SOWA Privacy Chat for your team

Business and Enterprise customers receive the application through their account contact. Tell us your team size and we will set it up.

Talk to sales See plans