Can Your Security Cameras Run AI Without Sending Images to the Cloud?
- Written by
- Agentinosaurus Team
- Published
- Reading time
- 3 min read
Local AI analyses camera frames on a computer at your home or business. Here’s how it compares with cloud AI, what hardware it needs, and what Agentinosaurus already keeps off its own servers today.
Can AI analyze camera images without uploading them?
Yes. Vision models small enough to run on an ordinary desktop computer can now look at camera frames and answer simple questions about them, such as whether someone is on the floor or a door is open. When the model runs locally, frames never leave your building for analysis.
Even without local AI, Agentinosaurus is designed so camera frames don’t pass through our servers. Station sends frames directly to the AI provider you choose, with your own key, and our servers only receive the answers.
| Cloud AI | Local AI (desktop Stations) | |
|---|---|---|
| Where frames are analysed | Your AI provider (TypeSafe AI or OpenAI) | The Station computer |
| Do frames reach Agentinosaurus servers? | No | No |
| Do frames leave your premises? | Yes, to the provider you chose | No |
| AI usage cost | Billed per check by the provider | None (your hardware and electricity) |
| Hardware | Any computer that runs Station | 8 GB memory or more; an NVIDIA GPU or Apple Silicon makes it faster |
| Works with phone Stations | Yes | No, desktop Stations only |
How does local AI work with Agentinosaurus?
Station downloads a verified open model once, runs it on the same computer and checks camera frames against your Events there. Your workspace still receives the results, so alerts, incidents and history work exactly as with a cloud provider.
Station installs the verified LiquidAI d1-2B Q8 model and runs a video test before enabling it. Check performance on your own hardware; mobile Stations currently use cloud AI.
Does local AI mean everything works offline?
No. Local AI keeps image analysis on your computer, but Station still needs the internet to receive your Events, deliver alerts to your phone and keep your account active. If the internet drops, analysis can’t be reported until the connection is back.
Local image processing and a fully offline system are different promises. Agentinosaurus makes the first.
Does Agentinosaurus receive your camera frames?
No. Here is what goes where with each setup.
- Camera frames: from the camera to Station on your network, then to your chosen AI provider (cloud) or nowhere (local).
- Hidden areas: blacked out on Station before any frame is analysed, stored as a snapshot or sent.
- Incident snapshots: kept on the Station computer for seven days, never uploaded to our servers.
- What we receive: Event results, incident details, usage counts and device health.
Can you use local AI for some cameras and cloud AI for others?
Yes. Each Station has a default AI provider and each camera can override it. A common split: keep a bedroom or treatment-room camera on a local model and the driveway on cloud AI.
Camera Chat, where you ask a camera a question, uses OpenAI separately. If the goal is to keep every image on your premises, use Event monitoring on a local model and don’t use Camera Chat for that camera.
What does local AI cost?
There is no per-check AI bill, so a camera that is analysed continuously costs the same as a quiet one. You still pay for your Agentinosaurus plan, and the computer, its electricity and the one-time model download (about 2 GB) are yours.
For comparison, cloud AI at default settings costs roughly $0.15–0.20 a day for a quiet camera that analyses on movement and about $2.50 a day for one analysed continuously. If you watch many cameras around the clock, a local GPU can pay for itself.
| Part | Recommended | Why |
|---|---|---|
| Memory | 8 GB or more | The model is loaded into memory |
| GPU | NVIDIA or Apple Silicon (optional) | Much faster checks |
| Disk | About 3 GB free | Model files and runtime |
| System | Windows, macOS or Linux desktop Station | Phone Stations use cloud providers |