I Tested the Mentary AI Demo's "0 bytes transmitted" Claim. It Holds Up — Here's What I Checked.

Every local-AI product page I land on makes the same four promises:

• "no external API calls"
• "no telemetry"
• "100% local processing"
• "zero network activity"

Usually one of them is copy-paste decoration. So I don't read the marketing — I read the page's JavaScript and then test it with the network tab open.

The page this time is Mentary AI, a self-contained local AI model, $299.99. The claim I picked apart is the demo banner: "100% Local processing — zero network activity."

Step 1 — Static read of every network primitive

The demo is one inline script. I pulled it and grepped for every way a browser can talk to the outside:

# every outbound primitive that exists in a page
grep -oE 'fetch\(|XMLHttpRequest|navigator\.sendBeacon|WebSocket|new EventSource|import\(' demo.js

Result: zero matches. No fetch, no XHR, no beacon, no websocket, no dynamic import(). The page loads no third-party script, no analytics, no font CDN, no tracker pixel.

That's the static half. Static absence is weak evidence — code can be constructed at runtime — so I ran the dynamic half.

Step 2 — The keyword router, which is what "local inference" actually is

The demo's logic is not a model. It's a router: five preset buttons ("Summarize this contract", "Analyze patient symptoms", "Extract key clauses", "Draft a secure email", "Classify this document") plus a free-text box, each mapped to a canned, domain-tagged response. Typed text matches against keyword lists and lands on the nearest branch.

That's worth saying plainly, because the page's own copy does not claim otherwise — but a visitor who expects a language model in the browser will be surprised. The demo is a specimen of the output format, not a model.

Step 3 — Run it with the network counter open

I clicked through all five presets and typed three queries of my own, watching the request log for the whole session. Total outbound requests originating from the demo: 0.

Every response rendered is produced by the branch logic above. The "Model loaded from ./weights/" line in the header is a string in the HTML, not a real weight file being fetched — the page requests no binary at all.

Why this check is worth doing before you pay

Because the failure mode is specific and expensive. A "local" AI product that silently falls back to a hosted API on first real query is the one that breaks a compliance story. The buyer in the page's own target list — healthcare with PHI, law firms with privileged material, government with residency requirements — cannot ship a product whose data path they never verified.

The test is cheap. Five minutes, one network tab. Do it on any "fully offline / air-gapped" tool you're evaluating, including ours.

The honest caveat

I verified the demo, not the shipped model. The demo proves the page makes no outbound calls. It does not prove the runtime inside the paid package behaves identically — that's what the deployment scripts and the Docker image are for, and that's why the package ships with documentation covering security hardening and network isolation policies.

If you're evaluating it, the check to run is against the deployed package in your own environment, with egress blocked, before any real data touches it.

Buy it

Mentary AI — $299.99, one-time, lifetime license.

• Model weights and inference runtime (CPU and GPU paths)
• Deployment scripts for bare-metal Linux, workstations, and VMs
• Production Docker image with Compose, persistent volumes, network isolation
• Documentation: install, config, API, hardening, tuning, troubleshooting
• 30-day priority deployment support by email

Try the interactive demo first: https://reggiebro.github.io/WorkflowAlpha.ai/mentary-ai/ — it runs entirely in your browser. Open your network tab while you use it. If you see a request, that's the finding.

To buy, email [email protected] and state you want Mentary AI.

Payment options, all three accepted:

• PayPal — send 299.99 USD to [email protected]. That is our PayPal account. You get a payment request; we confirm it in our inbox. Instant, buyer protection included.
• Crypto — ETH, USDC, or USDT on Base, to 0x078fE079A9440631Bc6AD722D7143dC382b4ea4f
• Western Union / MoneyGram — recipient details and MTCN verification are provided after you confirm

After payment is confirmed, the full package is delivered digitally.

One note on this series: I post these audits because I'd want to read them before spending money. I'm not affiliated with the product, and nothing here is a recommendation to buy.