The problem
Online manipulation doesn't happen in a single message
When a predator targets someone online, they rarely send one obviously dangerous message. Instead, they spend weeks or months building trust, creating emotional dependency, pushing boundaries a little further each time, and conditioning the victim to keep the relationship secret. By the time something clearly wrong happens, the victim is often already deeply trapped.
The same pattern plays out in financial scams, romance fraud, and sextortion — a slow build of trust, followed by escalating pressure. Today's safety tools almost always miss this because they look at individual messages in isolation, or they scan content on a server somewhere in the cloud. Neither approach can see the full picture.
Why current tools fall short
Three problems no one has solved — until now
They can't see encrypted messages. Apps like WhatsApp use end-to-end encryption, which means the message exists as readable text only on the sender's and receiver's screens — never on any server in between. Cloud-based safety tools are completely blind to these conversations, which now make up the majority of mobile messaging.
They can be bypassed with a single tap. Tools that monitor network traffic only work on Wi-Fi. The moment someone switches to cellular data, all monitoring disappears — no technical skill required.
They depend on app makers' permission. Many parental tools connect through the official APIs of social apps. When those apps change or revoke API access — which they increasingly do — the tool stops working with no fallback.
Agent Hita sidesteps all three problems. It reads what's displayed on the device screen — the same way a screen reader for the visually impaired does — so it works across every app, on any network, including end-to-end encrypted conversations, without needing permission from any platform.
How it works
AI that runs on your phone, not on our servers
When Agent Hita is active, it analyzes incoming messages in the apps you choose — WhatsApp, Instagram Direct, SMS, and others. As messages appear on screen, the app analyzes them using three layers of on-device intelligence:
First, it checks messages against thousands of known harmful phrases organized by type of harm — grooming language, financial pressure, threats, and more. Second, it scores individual words that often signal harmful intent even when they don't match a known phrase pattern. Third, a compact AI model running entirely on the device provides deeper understanding of context, tone, and meaning — catching manipulation expressed through indirect or coded language.
All three layers combine into a risk score. That score is compared against thresholds that automatically adjust based on who is being protected — the bar is set lower for children and teenagers, because the same message carries more risk when directed at a younger person.
The raw message text is never written to storage and never leaves the device. Only if a risk threshold is crossed does Agent Hita send anything — and what it sends is a structured alert containing only the category of risk (e.g., "grooming" or "financial scam") and the severity level. No message content. No contact names. No conversation history.
Privacy by design
What we store and send — and what we never do
On the device, the app stores only the risk category, severity score, app name, and an anonymized identifier for the contact — the contact's actual name is scrambled with a one-way mathematical function before it ever touches storage.
When a guardian alert is sent, it contains the risk category, severity level, and a recommendation for next steps (such as specific reporting resources for that type of harm). It never contains message text, the contact's name or number, or any conversation history.
Alert tiers
You choose how much information flows out
Local only: everything stays on the device. No alerts leave at all.
Metadata alert: a guardian receives the risk category, severity, and the name of the app where it occurred — nothing else.
Summary reporting: periodic summaries showing patterns across contacts and categories, without message content.
Evidence bundle: when explicitly enabled, structured metadata for law enforcement or child safety reporting.
Built-in protection against misuse
Designed so it can never be weaponized against the people it protects
Monitoring tools can be abused. An overbearing or abusive adult could theoretically use a safety app to surveil and control someone — which would make Agent Hita part of the harm rather than a solution to it.
We built a specific safeguard against this. Before Agent Hita analyzes any message, it checks whether the person being monitored is reaching out for help — searching for phrases like "am I being monitored," "I feel unsafe," or "coercive control." If it detects that, three things happen automatically: analysis stops, crisis resources and helpline numbers are shown privately to the person on the device, and no notification is sent to the guardian. The moment is never recorded anywhere.
Additionally, every 90 days, the app privately prompts the monitored person to confirm they still consent to monitoring. Their response is never shared with anyone.
Who it's for
Protection for anyone who needs it
Parents of children and teenagers. Agent Hita gives parents meaningful early warning about the patterns that precede serious harm — grooming, sextortion attempts, dangerous contacts — without turning into an intrusive surveillance tool that reads every message.
Guardians of vulnerable adults. Adults with cognitive disabilities, those in early-stage dementia, or people in isolated or controlling situations can benefit from the same protection, with settings calibrated for adult-level autonomy and dignity.
People protecting themselves. Adults who are experiencing harassment, coercive control, or financial pressure can run Agent Hita for their own benefit — receiving private alerts and safety guidance without involving anyone else.
Our commitment
Safety without surveillance
Agent Hita is designed around a clear belief: that protecting people from harm and respecting their privacy are not competing goals. Every design decision — reading the screen rather than intercepting the network, running AI on the device rather than the cloud, sending categories rather than content, protecting the monitored person from coercive misuse — reflects that belief.
We are building technology that gives families more safety, more dignity, and more peace of mind, without turning personal devices into surveillance tools.
Because privacy shouldn't require blind trust, Agent Hita is open-source. Anyone — researchers, journalists, parents, or security experts — can read the code and verify that it does exactly what we say. No hidden data collection. No backdoors. View the source code on GitHub →