Questions about privacy, how the app works, guardian alerts, battery life, and more — answered plainly.
Yes — but only in memory, only during analysis, and it never stores or transmits the text. Agent Hita reads what's visible on screen in order to analyse it. The moment analysis is complete, the raw text is discarded. It is never written to storage and never sent anywhere. What gets stored locally is the result: a risk category, a severity level, and the app name — not the messages themselves.
A guardian email contains only:
The email never contains message text, contact names, phone numbers, conversation history, or anything else identifying the specific person or conversation involved.
No message content is ever sent to the cloud. The only data that leaves the device is:
All detection and AI analysis happens locally, on your device.
Yes. Agent Hita's source code is publicly available on GitHub. Anyone — researchers, journalists, parents, or security experts — can read the code and verify that it does exactly what we say. There are no hidden data collection paths, no backdoors, and no obfuscated network calls. View the source code →
The app is free for individual use. Commercial use requires a license.
Currently supported: WhatsApp, WhatsApp Business, Instagram Direct, Google Messages, Samsung Messages, and the standard Android SMS app. Only 1-on-1 conversations are analysed — group chats are intentionally skipped.
Additional apps can be added via remote configuration without requiring a new app update.
Yes. Agent Hita reads what's visible on screen — the same way a screen reader does — rather than intercepting the network. End-to-end encryption doesn't affect what's displayed on the device screen, so it works fully on WhatsApp, Instagram DM, and any other encrypted messaging app.
Agent Hita uses Gemma — a compact AI language model by Google — to provide deeper semantic understanding of messages. It runs entirely on the device, requires no internet connection once installed, and message content is never sent to any server.
The first time you open the dashboard, Agent Hita offers to set up Gemma. After you accept Google's Gemma terms, the app downloads and installs the model (~900 MB) automatically in the background — no manual steps needed.
If you decline, Agent Hita still runs its full rule-based detection pipeline, which covers thousands of known harmful phrases and word patterns. Gemma enhances detection — especially for coded or indirect language — but is not required for the app to work.
Eight categories of harmful behaviour:
LOW: Weak signals detected. Stored locally for pattern tracking. No notification sent.
MEDIUM: Moderate signals. A private local notification appears on the device only. Nothing is sent to the guardian.
HIGH: Strong signals corroborated across multiple detection layers. A local notification fires and a guardian alert email is sent (if a guardian is configured).
Thresholds are automatically lower for children and teenagers, so the same message triggers a higher risk level for a younger person.
Yes. Agent Hita is not a hidden surveillance tool. The app requires explicit setup by the person on the device, including accepting terms, choosing a user category, and optionally configuring a guardian. The Android Accessibility Service permission — which the app needs to function — is visible in the device settings, and cannot be granted without the device owner's direct action.
Agent Hita is designed to be consensual. The person being protected is always aware the app is installed.
Before analysing any message, Agent Hita checks whether the monitored person is searching for help or trying to understand their monitoring situation — phrases like "am I being tracked," "I feel unsafe," or "domestic abuse help." If detected, three things happen automatically:
This moment is never recorded anywhere. It is designed to protect the monitored person from situations where the monitoring itself has become a tool of control. Read the full Trust Specification →
Guardian alerts are throttled to a maximum of one alert per contact per 24 hours. If multiple HIGH-risk events occur for the same contact within that window, they are accumulated and sent as a single aggregated alert when the window expires. This prevents alert fatigue while ensuring the guardian is informed.
Agent Hita is event-driven — it only runs when a messaging app is in the foreground and a new message appears. It does not run in the background continuously or poll for changes. Rule-based detection runs in milliseconds. Gemma inference takes a second or two per message but is only triggered when a new message appears on screen.
In typical use the battery impact is minimal. Some Android manufacturers (notably Samsung) aggressively kill background services. Agent Hita addresses this by running as a foreground service with a persistent notification, which also gives you visibility that it is active.
Not currently. Agent Hita relies on Android's Accessibility Service API, which has no equivalent on iOS. Apple's platform restrictions prevent third-party apps from reading the screen content of other apps, which is the foundation of how Agent Hita works. An iOS version would require a fundamentally different approach, and we are evaluating what's possible.
Android 8.0 (API level 26) or higher is required. The Gemma on-device AI model requires a device with at least 4 GB of RAM and works best on devices with 6 GB or more. The app runs fully without Gemma on any supported Android device.
A false positive is when Agent Hita flags a message as risky when it shouldn't have — for example, teenagers using hyperbolic slang like "I'll kill you at this game." These are expected in any detection system.
You can report a false positive directly from the event detail screen in the app. This helps improve the detection models. False positive reports contain only the risk category and severity — never the message text itself.
This is a real challenge. Teenagers frequently use words like "kill," "destroy," "attack," and "I'll end you" in gaming and social contexts. Agent Hita addresses this in several ways: