Roadmap

What we shipped, what we are building, and what we won't build.

Timings are estimates, not promises, and shipping order can change. Anything marked shipped is available today. Last reviewed 25 October 2026.

Shipped recently

Released and available today.

  1. Masks now use bullets, so ChatGPT's composer can no longer strip them, and restoring decoys in ChatGPT replies works again. Still local, with no new permissions.

  2. restoreJson puts real values back into tool-call arguments one value at a time and never restores keys. Strict mode refuses invented, expired or altered tokens.

  3. Flags secrets and personal data while you code, copies a scrubbed version for your assistant, and restores the real values when you paste the answer back.

  4. A small local model that finds names, organisations and places in plain sentences, for text where no label gives them away.

  5. n8n node and Gateway Guard, in beta

    2026

    Redaction nodes for n8n workflows, and a self-hosted OpenAI-compatible hop that tokenises messages and gates tool calls with a default-deny policy.

Building now

In development or testing. Shipping order can change.

  1. A prompt hook that stops a prompt carrying personal data or a secret before your coding agent sends it, and shows a clean version to send instead. Tested end to end in Claude Code and Codex; tested against Gemini CLI's documented hook contract.

  2. Licensed releases that add realistic decoys, your own watchlists and custom rules, with an opt-in seven-day trial and no card. Current store builds keep their legacy free features.

  3. African national ID numbers

    In development

    Nigeria's BVN and NIN, the Ghana Card and Kenyan national ID numbers in the identity pack, found when they appear next to a label.

Next

Planned for the libraries on the way to 1.0.

  1. Share token mappings across workers and processes instead of keeping them in one process's memory.

  2. Published benchmarks against Microsoft Presidio

    Planned

    Committed to the repository, so anyone can rerun them and check the numbers.

  3. Today's email detector matches ASCII addresses only.

  4. Steadier accuracy and a stable API for finding names, organisations and places in free text.

Later

Committed direction, without a date yet.

  1. The hosted gateway will use the same audit event schema as the library, so moving from local logging to hosted won't reshape your events. The library stays free and complete.

Not planned, and why

Requests we have decided against, so you can plan around them.

  1. Intercepting your network traffic

    Not planned

    A desktop proxy would need you to install a root certificate, the most powerful credential on your machine. Sether edits the text before it is sent instead.

  2. "Zero-knowledge" claims

    Not planned

    That label can't prove data didn't leave some other way. Sether's audit events record what was replaced and how long it was, never the original value.

  3. Guessing traits from a name

    Not planned

    Placeholders say what kind of detail was replaced, such as <NAME_1>. Sether will not infer gender, role or anything else you didn't write.

  4. Rewriting your prompt with another AI

    Not planned

    Sether finds and replaces specific details. It does not rephrase what you wrote, so nothing changes that you can't see.

  5. Shield runs in Chrome on a laptop or desktop. On a phone you can try the browser demo and send yourself the install link.

Want something on this list?

Open an issue on GitHub with the data or workflow you need protected. Before you rely on Sether, read what it does not do today.