PromptGuard is a lightweight, open-source web tool that removes personal and confidential information from any text in real time. Paste a log file, an email thread, a config file or a bug report on the left, and get a clean, AI-ready version on the right, with every sensitive item replaced by a clear placeholder tag such as [EMAIL_REDACTED] or [API_KEY_REDACTED].

Everything runs locally in your browser. Your text is never uploaded, stored or logged, and there is no server at all.

👉 https://xiaoyulei317-droid.github.io/PromptGuard---AI-Privacy-Sanitizer/

Click Load Sample Text to see every detection rule in action.

AI assistants have become part of everyday work, and so has the habit of pasting raw text into them. That text often contains more than you realize:

• 📧 Customer emails and phone numbers in support tickets and CRM exports
• 🌐 Internal IP addresses in server logs and stack traces
• 🔑 API keys and tokens in .env files, curl commands and config snippets
• 🏢 Client names, project code names and colleagues' names in meeting notes

Once that data is sent to a third-party AI service, you no longer control it. It may be stored, reviewed or used for training, depending on the provider and your plan. A single leaked API key can cost real money within minutes, and sharing customer data can break GDPR, CCPA or your company's policies.

PromptGuard adds a 2-second safety step between your clipboard and the AI. You still get useful answers from the AI, because the structure and meaning of your text stay intact. Only the sensitive values are removed.

The detectors are tuned so that ordinary text is left alone. These are not redacted:

• Dates and times: 2026-10-21, 14:30, 2026-10-21T14:30:00Z
• Version numbers and decimals: 3.14.2, 3.14159
• Money, ranges and IDs: $12,500, 5000-8000, #48213
• UUIDs, credit-card-length numbers and tables of numbers
• Code such as std::io, Add::new and Python slices like a[1::2]

When two rules overlap, the stronger one wins. For example, a keyword inside an email address never leaves part of the address visible.

• Live preview: the output updates as you type or paste
• Color-coded highlighting: each type of redaction has its own color, so you can review the result at a glance
• One-click copy of the clean text (with a fallback for older browsers)
• Load Sample Text button for a quick demo
• Drag and drop a .txt file straight into the input box

• Number of items redacted, broken down by type
• Character and word count (input vs. output)
• Estimated token count, so you know how much of the model's context window you will use

• Clean, modern design built with Tailwind CSS
• 🌗 Dark / light mode that follows your system setting and remembers your choice
• Fully responsive, from phones to wide screens
• Keyboard accessible, with ARIA switches and visible focus rings

• No backend, no analytics, no cookies, no tracking
• No web fonts or third-party requests, apart from the Tailwind CSS CDN
• Only your rule toggles and theme preference are saved, in localStorage. Your text and keyword list are never stored.
• Works offline once Tailwind has loaded. Add a prebuilt CSS file for fully offline use.

PromptGuard is a single index.html file. There is no build step and nothing to install.

Option 1: open it locally

git clone https://github.com/xiaoyulei317-droid/PromptGuard---AI-Privacy-Sanitizer.git
cd PromptGuard---AI-Privacy-Sanitizer

Then double-click index.html to open it in your browser.

Option 2: serve it locally

npx serve .

Option 3: deploy to GitHub Pages (free)

1. Push this repository to GitHub.
2. Go to Settings → Pages.
3. Under Source, choose Deploy from a branch, select main and / (root), then click Save.
4. After about a minute, your tool is live at https://xiaoyulei317-droid.github.io/PromptGuard---AI-Privacy-Sanitizer/.

It also works as-is on Netlify, Vercel, Cloudflare Pages or any static host.

Input

Reporter: Jane Doe <jane.doe@example.com>
Phone: +1 (555) 123-4567
Affected host: 192.168.1.42
OPENAI_API_KEY=sk-proj-••••••••••••••••••••••••
This belongs to Project Phoenix. Deployed on 2026-10-21.

Output (with the custom keywords Jane Doe and Project Phoenix)

Reporter: [KEYWORD_REDACTED] <[EMAIL_REDACTED]>
Phone: [PHONE_REDACTED]
Affected host: [IP_REDACTED]
OPENAI_API_KEY=[API_KEY_REDACTED]
This belongs to [KEYWORD_REDACTED]. Deployed on 2026-10-21.

All logic lives in two <script> blocks inside index.html:

1. Sanitizer engine (PromptGuard, pure JavaScript with no DOM access). Each rule runs its own regular expressions on the original text, in priority order: API Key → Email → IP → Phone → Keyword. Candidate matches pass through validation functions (for example isLikelyPhone() and isLikelyIPv6()) that filter out look-alikes. Matched character ranges are reserved, so overlapping matches can never split one another. The text is then rebuilt with placeholder tags.
2. UI layer. Handles input events, toggles, keyword chips, theme, clipboard and statistics.

Matching runs in linear time on typical input. A 1 MB document is processed in about 50 ms.

Token counts are a heuristic: about 4 characters per token for Latin-script text and about 1 token per Chinese, Japanese or Korean character. Real counts vary by model and tokenizer.

Want to detect a new kind of secret? Add a pattern to the API_KEY_PATTERNS array in index.html:

// Example: detect "acme_live_..." internal tokens
{ re: /(?<![\w-])acme_live_[A-Za-z0-9]{24,}/g },

Want different placeholder text? Edit the tag values in the RULES object:

email: { label: 'Email', tag: '<EMAIL>' },

PromptGuard is a safety net, not a guarantee. Detection is pattern-based, so:

• It does not recognize arbitrary names, addresses or other free-form personal data. Add those as Custom Keywords.
• Short local phone numbers without an area code (such as 555-1234) are intentionally skipped, because they look the same as number ranges.
• New or unusual API key formats may not be recognized yet. PRs are welcome.

Always review the output before sharing it.

• Numbered, consistent tags ([EMAIL_1], [EMAIL_2]) so the AI can tell entities apart
• Restore original values in the AI's reply (reversible mapping)
• More detectors: credit cards (Luhn-validated), IBANs, national ID numbers, street addresses
• Browser extension
• Optional encrypted saving of keyword lists

Contributions are welcome!

1. Fork the repository.
2. Create a feature branch: git checkout -b feature/new-detector
3. Commit your changes: git commit -m "Add detector for XYZ tokens"
4. Push the branch and open a Pull Request.

When you add or change a detector, please include examples of text it should catch and text it should not catch.

Released under the MIT License. Free to use, modify and distribute.