Sentinel runs a fine-tuned AI model entirely on your Mac. Your messages never leave your device. No API keys. No cloud. No internet required.
Free during beta. Your data stays on your Mac.
The Problem
The most dangerous texts aren't obvious scams - they're the ones from people you trust. Guilt trips disguised as concern. Threats wrapped in "lol." Lies that sound just plausible enough. You don't catch them because you're not supposed to.
A second pair of eyes that can't be gaslit.
If someone in your life makes you constantly question yourself, Sentinel can help you trust your instincts.
If you or someone you know is in crisis: National DV Hotline - 1-800-799-7233
Privacy First
Sentinel runs a purpose-built AI model fine-tuned specifically for deception detection, entirely on your Mac. No servers, no APIs, no internet connection needed. Your texts are never transmitted anywhere.
The AI model runs entirely on your Mac's hardware
Zero network calls. Works in airplane mode
No accounts, no tokens, no third-party services
Messages never leave your device. Ever
How It Works
Sentinel sits in your menu bar and works silently in the background. No windows to manage, no workflows to learn.
Download the app and drop it in your Applications folder. A purpose-built AI model installs locally on your Mac. Takes about 2 minutes.
The on-device model silently evaluates every incoming message against an 8-step behavioral rubric drawn from deception research. All processing happens locally.
When a threat is detected, you get an immediate notification with context on what was flagged, the specific category, and why it was suspicious.
What It Detects
Trained on thousands of labeled examples spanning the full spectrum of manipulative messaging.
Fake identity claims, credential theft attempts, impersonation, and social engineering tactics designed to steal information.
Guilt-tripping, gaslighting, love bombing, and other tactics used to control through exploiting emotions and attachment.
Investment scams, romance scams, advance-fee fraud, and deceptive financial requests disguised as opportunities.
False urgency, fake scarcity, ultimatums, peer pressure, and coercive language meant to force quick decisions.
False alibis, selective withholding of information, blame-shifting, and strategic distortion of facts.
Threats and intimidation hidden behind humor, sarcasm markers like "lol" and "jk," or plausible deniability.
Built on Research
Sentinel's detection engine is built on deception research from UC Berkeley and Oxford. The model uses an 8-step behavioral rubric that maps to how manipulation actually works in real conversations.
Unlike generic AI classifiers, Sentinel was fine-tuned on thousands of carefully labeled examples that span the full taxonomy of deceptive messaging - from phishing to emotional coercion to masked threats.
This isn't a toy. It's a research-backed detection engine built by Jonathan Bodea, designed to catch what humans miss in the messages they trust most.
Technical Specs
A fine-tuned model small enough to run on your laptop, powerful enough to catch what you'd miss.
FAQ
Sentinel only analyzes incoming messages from other people - it never reads messages you send. Analysis happens entirely on your Mac using a locally installed AI model. No text is ever transmitted anywhere.
Yes. After the initial setup (which downloads the AI model once), Sentinel works completely offline. No internet connection needed.
Sentinel runs on any Mac with Apple Silicon (M1, M2, M3, M4, or later) running macOS 13 or newer. It uses about 4GB of RAM while running.
No. Sentinel was specifically trained to distinguish sarcasm, jokes, and genuine emotional expression from actual manipulation. “You’re literally the worst lmao” won’t trigger an alert. “You’re the worst, and if you don’t do what I say, I’ll make sure everyone knows it” will.
Spam filters catch mass-sent junk. Sentinel catches targeted, personal manipulation - guilt trips from people you know, threats disguised as jokes, scams that use your name and context. These are the messages spam filters are designed to ignore.
No. The AI model comes pre-trained. Your messages are never used for training, never stored beyond the analysis moment, and never leave your device.
Sentinel achieves 90% precision with a 12.8% false positive rate on our benchmark. It’s trained on thousands of labeled examples using an 8-step behavioral rubric developed from academic deception research.
Sentinel is free during the beta period. We’ll announce pricing before any changes.
Early Access
Sign up for early access and be the first to get on-device AI protection for your messages. No spam, no data collection - just a launch notification.