Private AI Text Assistant

What can the Private AI Text Assistant do?

It is a writing and review workspace for text you would rather not send to a hosted chatbot. Paste an email draft, a caption, OCR output, a policy excerpt, or a prompt you are about to reuse. You can scan it for obvious identifiers, make a redacted copy, summarize it, rewrite it, or pull out structured facts.

The page is usable before any model download. Privacy Scan’s first pass is a local pattern matcher that inspects the entire paste. The optional MiniCPM5-2B engine is for deeper reading, not a required first step.

How local AI protects your text

Your text stays on this device. The site can still load ordinary page assets, and it can download model weights when you click Load local AI. That download is not the same as sending a prompt to an API.

This tool does not call CreatorPrivacyKit for inference, does not use Hugging Face inference endpoints for your writing, and does not fall back to OpenAI, Anthropic, or Gemini. If WebGPU is missing, the model path is disabled. The scanner still runs.

Privacy Scan

Privacy Scan looks for shapes that often leak: emails, phone-like numbers, IPv4 and IPv6, URLs, JWT-like strings, bearer tokens, common API key prefixes, private key blocks, and similar assignments. That pass does not need the 1.8 GB model, and it is not clipped to 32,000 characters. A secret at character 90,000 is still in scope. Create Safe Version redacts those spans across the full document.

A deeper local AI scan can add names, client context, and other phrasing the regex will miss. Long documents are reviewed in bounded sequential chunks rather than loading the model’s theoretical 131K context into browser GPU memory. Neither layer is complete. Do not treat a clean result as proof that a document is safe to publish.

Local AI summarizer

Summarize is for long pastes: notes, briefs, transcripts, policies. Choose a short paragraph, a longer recap, bullets, action items, or key facts. Practical MiniCPM input is about 8,000 tokens on a typical desktop GPU — far below the model’s theoretical 131K context, which this page does not promise in a browser. Longer documents can still be privacy-scanned in full.

Anonymous text rewriter

Anonymize rewrites while keeping meaning. Light mode targets obvious identifiers. Standard and strict modes ask the local model to generalize client, workplace, and location detail. Typical replacements look like John Smith → [NAME], an inbox → [EMAIL], a street address → [ADDRESS]. The original stays in the left pane; the safe version is copied from the right.

How it works

Input stays in the page. Privacy Scan runs as JavaScript in this tab. If you load the model, Transformers.js fetches the ONNX weights into browser cache and runs MiniCPM5-2B on WebGPU inside a dedicated worker. Generation can be stopped. The main UI thread is not the inference loop.

Other local tools can pass extracted text through same-origin session storage. Document contents are not placed in query strings.

Browser requirements

For the local model: a recent desktop Chrome or Edge, WebGPU, and shader-f16. That combination is what the q4f16 build needs. Firefox, Safari, and most phones may run the scanner and still fail the model. This page will not silently switch to a cloud LLM.