Live demo

The actual screenpipe redaction models, running entirely in your browser via WebAssembly — your text and images never leave your machine (that's the whole point: local-first PII redaction). First use downloads the model once (~280 MB text, ~110 MB image), then it's cached.

Text redactor v45 · 278 MB INT8

One captured fragment per line — window titles, terminal output, OCR, chat (exactly how screenpipe redacts each string as it's captured). Catches API keys, passwords, connection strings, emails, repos…

Terminal | export OPENAI_API_KEY=sk-proj-Ab12Cd34Ef56Gh78 psql | DATABASE_URL=postgres://admin:S3cretPass@db.acme.com/prod Zoom | calendar invite to daniel.okafor@northwind.io Arc | PR opened on github.com/acmeco/widget-engine

redact:

Image redactor rfdetr_v11 · 109 MB

Finds and blacks out PII regions in a screenshot — names, IDs, addresses, secrets and more. Pick a sample or upload your own. Works best on clean, standard app UIs; unusual or low-quality screens may be missed or over-boxed.

sample: Upload

Zero-leak rate — local models vs frontier & cloud

Text PIIdesktop telemetry strings

Gemini 3.1 Pro	91.0%
GPT-5.5	90.7%
Claude Opus 4.7	87.8%
pii-redactor · local	86.7%
Google Cloud DLP	37.7%
Microsoft Presidio	35.4%

Image PII regionsIoU ≥ 0.30

pii-image-redactor · local	98.9%
Gemini 3.1 Pro	4.2%
GPT-5.5	3.2%
Google Cloud DLP	2.6%
Claude Opus 4.7	2.1%
Microsoft Presidio	0.5%

Zero-leak = share of items where every PII span (text) or region (image) is caught. Local models run fully offline (~10 ms text · ~120 ms image). Full methodology, confidence intervals & per-framework breakdowns in the leaderboard.

Runs entirely in your browser via transformers.js (text) and onnxruntime-web (image) — nothing is uploaded. Models: pii-redactor · pii-image-redactor. Synthetic samples only — no real PII.