Personal data and secrets become placeholders before your text leaves your machine — and the real values come back only into your local tools.
How it works
A fine-tuned GLiNER model and a set of pattern recognizers find the values, and an in-memory vault swaps each one for a placeholder and back. The model you call sees only placeholders; the real values return only into the local tools you allow — never into a shell command.
What you get
A fine-tuned GLiNER model plus a set of pattern recognizers, on your CPU or GPU. Your text goes nowhere, and the vault lives in memory only.
Masks tool output — files, logs, command output — not only prompts, and keeps links to code hosts and documentation readable. Built and tested with Claude Code, through the mod or the gateway.
Mask before your LLM call and restore after it. Or point your client at the gateway, in front of the Anthropic Messages API or the OpenAI Chat Completions API.
One JSON file decides what gets masked, for the library, the CLI, the gateway and the Claude Code mod: per category, mask it, ask first, or leave it readable.
A placeholder belongs to its session, and one value always gets the same placeholder. Streaming restore puts the real values back into chunked replies.
Benchmarked on coding-agent traffic and held-out document sets, with the cases it loses published too. See the benchmark.
