Local-first: camera AI that stays at home
- Written by
- Agentinosaurus Team
- Published
- Reading time
- 4 min read
Desktop Stations now analyse your cameras with LiquidAI d1, on your own computer, by default. No API key to manage, no usage bill and no frame sent to a cloud AI provider. Test results hold their own, and setup takes a click or two.
Local-first, by default
When Agentinosaurus launched, every Agent needed a cloud AI provider to think. Your Station sent frames straight to OpenAI or TypeSafe AI with your own API key, never through our servers, but always to someone.
Local-first changes the default. A desktop Station now runs LiquidAI d1-2B on the computer next to your cameras and analyses every check right there. Cloud providers are still one switch away, per camera. They’re just no longer where you have to start.
No API key to manage
A cloud provider comes with homework: create an account, add a payment method, generate a key, paste it in, assign it to a Station, keep it safe and replace it when it changes. If the key is revoked or the account runs out of credit, your cameras stop analysing until someone notices.
With a local model there is no key at all. There’s nothing to create, nothing to rotate, no credential stored anywhere and no card that can be declined at 3 a.m. The model is a file on your computer, and it keeps working as long as the computer does.
No usage bill
Cloud providers bill every check by the token. That makes sense for a camera that looks when something moves, and it adds up fast for one that watches all day.
A local model has no provider bill. We still record usage, at $0, so you can see how busy each camera is. A camera that analyses around the clock costs the same as one that checks twice a day: the electricity of a computer that’s on anyway. That makes continuous analysis an easy yes for the rooms where it matters, instead of a line on next month’s invoice. Your Agentinosaurus plan and camera slots are billed as usual.
Private by design
With a local model, camera frames are analysed on the computer they were captured on. They’re never sent to a cloud AI provider and never pass through our servers. Station doesn’t fall back to the cloud when something goes wrong; a camera that can’t be analysed locally tells you so instead.
Everything that already protects your footage still applies. Hidden areas are blacked out and faces are blurred before the model sees a frame. Incident snapshots stay on your Station. Your workspace receives only the result, which is what powers your Messages, incidents, Push notifications and Webhooks.
Station still needs the internet to receive your Events and deliver your alerts, and Camera Chat, where you ask a camera a question, still uses your OpenAI key.
Results that hold their own
Small and private doesn’t have to mean worse. In our October 28 comparison, all three models received the same frames, timestamps and Events from warehouse, fall and road-accident footage. On the checks with reference labels, d1 matched or beat OpenAI on every measure. TypeSafe AI remained the most accurate overall.
| Measurement | OpenAI | TypeSafe AI | Local d1-2B |
|---|---|---|---|
| Agreement with labels | 71.4% | 92.9% | 75.0% |
| Events caught (recall) | 50.0% | 80.0% | 50.0% |
| False alarms on negative checks | 5.6% | 0.0% | 5.6% |
| Falls and ordinary activity | 100% | 100% | 100% |
| Warehouse | 72.7% | 100% | 81.8% |
| Road accidents | 58.3% | 83.3% | 58.3% |
Runs on everyday computers
d1 is a 2-billion-parameter model built for one job: deciding whether your Events are happening. That keeps it small: about a 2 GB download, and no data-centre hardware to run it.
That fits the computer you’d leave on anyway: an office PC, a gaming laptop or an Apple Silicon Mac, on Windows, macOS or Linux. Allow at least 8 GB of memory and about 3 GB of free disk space. An NVIDIA graphics card or Apple Silicon makes it quick: in our research test the model answered in a median of 0.98 seconds on a laptop RTX 4080, and in Station a complete check with two Events took about 2.4 seconds. On the processor alone it still works, more slowly, which suits one or two cameras that analyse when something moves.
Set up in a click or two
There’s nothing to configure. On Windows, leave Set up the local AI model checked in the installer. On a Mac, click Allow when you pair it. On Linux, press Enter when the pairing command asks. That approval happens on the computer itself, because our servers can never start a download there on their own.
Station does the rest in the background. It uses model files it already finds on the computer or downloads them, checks every file against its published SHA-256 checksum, picks the graphics card or the processor, and runs a short video test. Add your cameras in the meantime: they start analysing the moment the model is ready.
When a cloud provider still makes sense
Local-first isn’t local-only. Choose a cloud provider for a camera when you want the highest accuracy we’ve measured so far, when the Station is a phone, or when there’s no computer that can stay on. You can mix both on one Station: local for the living room, TypeSafe AI for the loading dock.
Availability
Local-first is available today in the latest Station for Windows, macOS and Linux, and LiquidAI d1-2B is the default AI on desktop Stations. Stations on iPhone and Android keep using a cloud provider.