State-of-the-art local AI. Runs entirely on your Mac. Prompts never leave the device. Works offline, behind firewalls, under NDA. No cloud. No account. No subscription. No telemetry. GDPR-safe by construction.
Nothing leaves your Mac. Enforced by the sandbox, not by policy.
Zero network paths in the binary. No URLSession. No sockets. No SDKs. Drop Wi-Fi mid-sentence. Keep typing.
Prompts, attachments, replies — on your Mac. Nothing sent to a vendor. Nothing cached on a server. Nothing used to train a model. If your NDA prohibits disclosure to third-party AI or cloud services, Airplane AI removes the disclosure path.*
No personal data transmitted. No third-party processor. Article 28 doesn't apply when no processor exists. Your SwiftData store lives in your user container — and stays there.
One entitlement: com.apple.security.app-sandbox. Nothing else. No network. No file access outside the container. CI fails the build if URLSession, NWConnection, WKWebView, or any analytics SDK appears in the binary.
€29.99 once. Mac App Store. No subscription. No per-seat pricing. No kill-switch. Keeps working on macOS 15+.
State-of-the-art local AI. Metal-accelerated on Apple Silicon. Sub-second first token after warmup. Open weights. Auditable. What model, and why →
* NDA-safe for the general case — typical NDAs prohibit disclosure to third-party AI or cloud vendors. Airplane AI removes that path: no vendor, no disclosure. Unusual clauses exist. Read your specific NDA. We can't read it for you.
Cloud AI: offline. Airplane AI: keeps going.
Upgrade, wait, or log in again. Airplane AI has no limit.
NDA. Medical. Legal. Airplane AI never leaves your Mac.
Real screenshots. Every token generated on your Mac.
| Airplane AI | ChatGPT | Claude | Gemini | |
|---|---|---|---|---|
| Works offline / airplane mode | ✓ | ✗ | ✗ | ✗ |
| Data stays on your Mac | ✓ | ✗ | ✗ | ✗ |
| No account required | ✓ | ✗ | ✗ | ✗ |
| No subscription | ✓ | ✗ | ✗ | ✗ |
| GDPR-compliant by construction | ✓ | ✗ | ✗ | ✗ |
| No training on your data | ✓ | ✗ | ✗ | ✗ |
| Works behind a firewall | ✓ | ✗ | ✗ | ✗ |
| No rate limits | ✓ | ✗ | ✗ | ✗ |
| Works during outages | ✓ | ✗ | ✗ | ✗ |
| On-device speech | ✓ | ✗ | ✗ | ✗ |
No API keys. No sign-up. No configuration.
Google's open-weight Gemma 3n E4B. ~4B-parameter quality. Quantized to Q4_K_M — four-bit compression that preserves reasoning and fits 16 GB of RAM with headroom for your real work.
Why Gemma 3n. Open weights. State-of-the-art at this size. Multilingual. Handles code. Follows instructions. Runs under llama.cpp with Metal. No vendor lock-in — weights are public and auditable.
Why Q4_K_M. Four-bit K-means quantization. Near-full reasoning quality at one-quarter the RAM. The sweet spot for 16–24 GB Macs.
URLSession, NWConnection, or any analytics SDK appears.App Sandbox. Zero network entitlements. No telemetry. No analytics. No crash reporting. Verified by CI at every build.