Mentary AI: Verifying a Private Inference Deployment on a Fresh Host

A deployment-runbook angle. I'm not a Mentary AI user yet — this is what the $299.99 package promises to install, and the exact order I'd verify each promise on a bare host before calling the deployment done.

The claim on the page

Mentary AI is sold as a fully self-contained local model: weights, inference engine, and runtime dependencies bundled, no external API calls, no telemetry, working in air-gapped environments. The license includes deployment engineering support — bare-metal, Docker, or integration into existing on-prem systems — plus 30-day priority support after purchase.

Every one of those claims is testable on a fresh host. Here is the order I would run them in.

Step 1 — The network isolation test (run this first)

This is the only claim that matters for the buyer profiles the page names — healthcare PHI, law firm privilege, government data residency. Everything else is performance.

# Block all outbound traffic for this process, then run an inference.
sudo unshare -n --map-root-user bash -c '
  ip link set lo up
  /opt/mentary/bin/mentary-serve --port 8080 &
  sleep 3
  curl -s http://127.0.0.1:8080/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d "{\"prompt\": \"Summarize this contract\"}"
'

If a response comes back, the model is genuinely local. If it hangs or errors on DNS, you have a hidden call out — which is the single thing that would disqualify the product for a privacy buyer. The page's own demo makes the same promise in the browser ("0 bytes transmitted"), and the page is honest about it: no discount, no urgency, just the price and the payment methods.

Step 2 — Confirm what is actually in the package

Before any tuning, inventory it. The page lists: model weights, the optimized inference engine, deployment scripts, a production Docker image with Compose orchestration, documentation, and 30-day priority support.

du -sh /opt/mentary/weights /opt/mentary/runtime 2>/dev/null
ls /opt/mentary/deploy/                       # bare-metal, VM, container scripts
docker images | grep -i mentary

A weight directory you can size and a runtime directory you can hash are what make the "no cloud dependency" claim auditable. If the model is pulled from a remote registry at first run, that is a finding, not a deployment detail.

Step 3 — Health check, then load check

curl -s http://127.0.0.1:8080/healthz
curl -s -w '\n%{time_total}\n' http://127.0.0.1:8080/v1/chat/completions -d '{"prompt":"Extract key clauses"}'

The page says CPU and GPU-accelerated runtimes are both included. That means two different performance profiles, and you should measure the one you will actually run. A CPU-only box that meets latency on a single concurrent user may fall over at five — that is the number to establish before anyone signs off.

Step 4 — The compliance mapping (why the price is $299.99)

The page states the architecture "simplifies compliance with HIPAA, GDPR, and data residency regulations." Read that carefully: it simplifies compliance, it does not confer it. What you actually get is the absence of a third-party processor in the data path — no logs sent externally, no usage analytics, no third-party access. That is a meaningful reduction in your vendor-assessment surface, and for many teams it is the whole reason to buy local.

What you do not get is a BAA, a SOC 2 report, or a signed residency guarantee. Those come from the vendor's own paperwork, not from a Docker image. Ask for them in writing during the 30-day support window — that is what the support period is for.

Buying

Live page: https://reggiebro.github.io/WorkflowAlpha.ai/mentary-ai/

The page states the price plainly at $299.99 USD, one-time, lifetime license, no recurring fees. Payment is PayPal to [e-mail] (buyer writes a note, seller confirms in the inbox), crypto at 0x078fE079A9440631Bc6AD722D7143dC382b4ea4f on Base — ETH, USDC, or USDT — or Western Union / MoneyGram. The page's contact address is [e-mail]. No discount is being offered anywhere on the page, and none is implied here.

Bottom line

For a team that cannot put prompts on someone else's hardware, a locally-bundled model is the correct architecture and the test is trivially cheap — unshare -n, then ask it a question. If it answers, the claim holds and the $299.99 is a reasonable price for removing a processor from your compliance scope. If it does not, that is worth knowing before the money moves, not after.

You can try the interactive demo on the page itself first — it runs entirely in your browser and reports zero bytes transmitted. Test before you pay.