Hi everyone, I’m Godfrey, building Sether. The question behind it is simple: does the model need the actual customer email to help write a better support reply? Often, it needs the task and context, while the exact identifier can stay with the application. Sether detects supported sensitive values, replaces them with stable tokens before the request goes out, and can restore recognized tokens from the same private mapping afterward. The library is open source under MIT. There are real limits: detection depends on the configured detectors and input format. A token the model changes may not restore. This is one layer in an application’s data handling, not a promise to catch everything or a compliance certification. The hosted gateway is still coming soon. I’d love feedback from developers working on support assistants or other text-based AI workflows. What would you need to test before putting a redaction step in your application? Please share fictional examples only.

Sether

Open-source redaction before your AI requests leave your app

Open-source redaction before your AI requests leave your app