AIgis

Secure your AI interactions.
A browser extension that masks secrets and optimizes tokens in prompts locally before sending them to LLMs.

---

What is AIgis?

AIgis (pronounced like Aegis, the mythological shield) is a "Client-Side Privacy Firewall" for Large Language Models.

Many companies and individuals hesitate to use tools like ChatGPT, Claude, Gemini, DeepSeek, or Grok due to data privacy concerns ("Shadow AI"). AIgis solves this by intercepting your prompt directly in the browser. It sanitizes sensitive data (PII) and optimizes data structures to save tokens before the request ever leaves your device.

Your data stays yours. The LLM provider only sees what you want them to see.

Features

🔒 1. Privacy & Sanitization

• PII Detection: Automatically detects and masks emails, phone numbers, IBANs, file paths, URLs, IP addresses, secrets and more.
• Smart Developer Mode: Switch between 'Strict' mode (masks everything) and 'Developer' mode (smartly ignores safe programming variables like localhost, 10.x.x.x private network ranges, or standard paths like ./node_modules/ to avoid breaking system code inputs).
• Custom Dictionaries & Regex: Define your own "forbidden words" (e.g., internal project names like Project Apollo) that get replaced with placeholders ([CUSTOM_1]). You can also specify powerful custom regex patterns directly in the UI by wrapping your regex in forward slashes (e.g., /EMP-\d{4}/). The flags i, m, s and u are respected per pattern — without flags, matching is case-insensitive by default (safer for masking); write any flag set without i (e.g. /Foo/m) to match case-sensitively. Invalid or ReDoS-prone patterns (like (a+)+) are rejected when you add them.
• Local Processing: All logic runs completely native within your browser. No data is sent to any 3rd party server for verification.

⚡ 2. Token Optimization (TOON)

• JSON to TOON: AIgis automatically detects bulky JSON inputs and converts them into TOON (Token-Oriented Object Notation) before sending.
• 30-60% Savings: TOON strips away redundant syntax (brackets, quotes) to drastically reduce token usage and API costs without losing structural integrity.
• Lossless & Reversible: Fully round-trip capable. Data converted to TOON is semantically identical to the original JSON, ensuring the LLM understands it perfectly.

🔄 3. Restoration & Utilities

• Context Restoration: AIgis seamlessly re-injects your original string data directly into the LLM's response block in real-time, meaning you read the real data while the LLM only ever saw the generic placeholders.
• Interactive Badges: AIgis renders intercepted targets as secure, clinical badges featuring the placeholder name (e.g., EMAIL_1). Hovering your mouse over the badge temporarily unmasks it to reveal your original, real data. A native Single Click immediately copies the raw, original data straight to your clipboard!
• Peek Mode: Need to verify multiple unmasked values across the entire page simultaneously? You can suspend the privacy overlays to visually expose all underlying real data! For a fast, temporary glimpse, simply hold down the \ (Backquote/Tilde) key (typically next to the 1). Releasing it securely snaps the placeholders back into place. For persistent unmasking while reviewing long outputs, toggle the "Peek" switch inside the AIgis Dashboard!
• Smart Copy: Don't worry about trying to extract information! Just highlight the LLM's text and copy it normally (Ctrl+C or Cmd+C). AIgis seamlessly intercepts the system's copy pipeline, unmasks the placeholders, natively decodes any TOON blocks to JSON, and injects the raw, safe data straight to your clipboard!
• Context Menus: Features a built-in Right-Click -> Decode TOON to Clipboard tool! If the LLM generates raw TOON syntax and misses the auto-decoder, simply highlight the block and right-click to natively restore the JSON to your clipboard!

🔔 4. PII & Config Subscriptions (Teams & Enterprise)

• Remote Rulesets: Subscribe to HTTPS-hosted JSON feeds containing shared custom wordlists and configuration templates — ideal for teams enforcing common masking rules. See subscription-sample.json for the fully documented feed format.
• Strictly One-Way: AIgis only pulls feeds (plain GET, no credentials, no local data ever attached). Feeds are cached locally, refreshed daily (with ETag support and offline fallback), and can be updated manually anytime.
• Managed Settings with Consent: A feed may provide settings overrides ("Apply Settings"). These are applied as a non-destructive overlay — your own settings stay saved and return when you unsubscribe. Every config (and every remote config change) must be explicitly confirmed before it takes effect, and managed settings are clearly badged with 🔒 Managed by [Feed] in the dashboard.
• ReDoS-Hardened Regex: Feed regex patterns are only executed if you mark the source as trusted — and even then each pattern passes a validation pipeline (compile check, length cap, catastrophic-backtracking heuristic). Regex entries from untrusted feeds stay inert (never executed or matched) until you enable them. Rejected patterns are surfaced in the UI.
• Unified Dictionary View: Subscription words appear alongside your local custom words (badged, read-only) with source filtering. Deleting a subscription optionally imports its words into your local list.

🛡️ 5. Secure Vault & Local Storage

• Local-First Architecture: AIgis operates without external servers. Your secure mapping Vault (which links generic Placeholders back to their real PII strings) is stored strictly and persistently inside your browser's native local storage.
• Auto-Pruning & Expiration: The Vault features a configurable auto-pruning mechanism. Mappings are assigned an automatic time-to-live (default 30 days) ensuring that old, stale context data is silently and permanently destroyed to reduce forensic risk. You can effortlessly track and individually renew or change these expirations through the Dashboard.
• Dual-Storage Security Model: AIgis utilizes a split-storage architecture for maximum security and convenience. General preferences, module toggles, and Custom Dictionaries sync across your personal devices via Google's secure sync storage. However, the actual sensitive data and PII Vault mappings are strictly quarantined to your machine's physical local storage.
• Interactive Dashboard: Use the built-in AIgis Dashboard to live-search and dynamically sort your protected Vault entries, trace global token-saving metrics, manage your Custom Dictionaries, and import/export your configuration files securely.

---

🌐 Supported Platforms

AIgis intelligently observes and injects its UI specifically onto the following LLM interfaces out-of-the-box:

• ChatGPT (chatgpt.com)
• Claude (claude.ai)
• Gemini (gemini.google.com)
• DeepSeek (deepseek.com)
• Microsoft Copilot (copilot.microsoft.com)
• Perplexity (perplexity.ai)
• Mistral (chat.mistral.ai)
• Grok (grok.com)
• Kimi (kimi.com)
• Xprivo (xprivo.com)

---

🛠️ How it Works

AIgis operates as a browser extension injecting a content script into supported LLM interfaces.

sequenceDiagram
    autonumber
    actor U as User
    
    box Your Browser (Secure Local Environment)
        participant UI as 🖥️ Web UI
        participant A as 🛡️ AIgis Engine
        participant V as 🗄️ Local Vault
    end
    
    box Internet (Untrusted)
        participant L as ☁️ LLM Provider
    end

    U->>UI: Submits Prompt ("Email foo@bar.com" + JSON)
    UI->>A: Intercepts request
    
    A->>A: 🔍 Scans text for PII & JSON
    A->>V: 🔐 Saves mapping (foo@bar.com)
    V-->>A: 🎫 Generates Placeholder [EMAIL_1]
    
    A->>A: 🎭 Masks PII & Compresses JSON to TOON
    
    A->>L: 🌐 Sends Request ("Email [EMAIL_1]" + TOON)
    Note over L: Processes data without<br/>seeing PII or bulky JSON
    
    L-->>UI: 💬 Streams Response ("Sent to [EMAIL_1]" + TOON)
    
    A->>UI: ⚡ Scans DOM & Detects Placeholder / TOON
    A->>V: 🔓 Looks up [EMAIL_1]
    V-->>A: 📦 Returns "foo@bar.com"
    A->>A: 🧩 Decodes TOON back to JSON
    
    A->>UI: ✨ Re-injects original data (Badges & JSON Blocks)
    UI-->>U: User reads the real, formatted text

Loading

---

Installation

AIgis works on all Chromium-based browsers (Google Chrome, Microsoft Edge, Brave, Opera, Vivaldi, Arc, etc.).

Option 1: Chrome Web Store (Recommended)

The easiest and safest way to install AIgis is directly from the official Chrome Web Store. Extensions installed this way receive automatic background updates. (Note: Due to Google's security review process for privacy extensions, Web Store versions may trail behind the latest GitHub releases by a few days).

Option 2: Developer Mode (Manual Installation)

If you want to test beta releases or install a specific older version, you can manually sideload the extension:

1. Download: Go to the Releases Page and download the latest AIgis-vX.X.X.zip (found under "Assets").
2. Unzip: Extract the ZIP file into a folder of your choice.
3. Open Extensions Page:
   • Chrome / Brave / Opera: Navigate to chrome://extensions/
   • Microsoft Edge: Navigate to edge://extensions/
4. Enable Developer Mode: Toggle the switch "Developer mode" (usually in the top-right corner or left sidebar).
5. Load Extension: Click the "Load unpacked" button.
6. Select: Select the folder you just extracted.

Ready! AIgis is now active. I recommend pinning the extension icon to your toolbar for quick access to the toggle switch.

💻 For Developers (Build from Source)

If you want to contribute or audit the code:

# 1. Clone the repo
git clone https://github.com/Karaatin/AIgis.git

# 2. Install dependencies
npm install

# 3. Build the project
npm run build

# 4. Load the 'dist' folder in Chrome/Edge (via "Load unpacked")

---

⚠️ Disclaimer

AIgis is provided "AS IS", without warranty of any kind. While this extension uses pattern matching and dictionaries to mask sensitive data, no regular expression or heuristic is 100% foolproof. Edge cases, typos, or obfuscated formats may bypass the detection engine.

You are ultimately responsible for the data you enter into LLM platforms. The developers of AIgis are not liable for any accidental data leaks, privacy breaches, or damages resulting from the use or failure of this software. Always double-check your prompts if they contain highly sensitive or classified information.