On-device video intelligence · Next-gen CCTV
Cameras that memorize everything they see — and let your AI agents read it.
The device turns cameras into more than a transcript — a continuous, plain-language record and a searchable index of what it means, produced by vision-language models running on optimized on-device inference. People can read and search it from a phone or browser, and the same intelligence is exposed to AI agents over MCP, so they can monitor, reason, and recall what happened in the physical world. Everything runs on the premises. Nothing is sent to the cloud.
The problem
The places with the most valuable video can't use cloud video AI. Factories, utilities, data centers, and secure facilities are barred — by policy or by air-gap — from sending footage off-site. For them, "nothing leaves the building" isn't a preference. It's a requirement.
And where the cloud is allowed, it's expensive. Cloud video AI bills by the camera and by the month — egress, storage, and the heavy cost of streaming every frame through a model — so watching more of your site only makes the bill grow.
At the same time, the AI agents organizations are starting to rely on are blind to the physical world. They can read your documents and call your APIs — but they have no grounded memory of what actually happened on-site. The device gives them that memory, without ever moving the footage.
How it works
It transcribes everything, continuously.
Every few seconds of video becomes a timestamped line in a growing transcript — a sentence describing what happened, in ordinary language. Each moment is also captured as structured data — objects, attributes, and tags — so the record can be filtered and searched with precision, not just read. A readable account of the day, written as it unfolds. See a real 3-hour transcript →
It's searchable by meaning, in plain English.
Ask for "someone entering without a hard hat" or "a vehicle at the dock after 6pm" and get the moments back at once. This is semantic search, not keyword matching: each frame — the image and its caption alike — is embedded into the same vector space as text the instant it's captured, so your plain-English query matches directly on meaning. No upload, no overnight indexing step — search is ready the moment something happens. Anyone you authorize can query it from a phone or browser on your network.
It runs on the premises. Nothing leaves.
Transcription, search, and alerting all run on a small appliance on your own network — offline-capable, no cloud bill, no per-event fees. If footage never leaves the building, the threat model collapses to local access, which you already control.
See it run
Watch the device work — on real footage.
No mockups. See the live demo, then read the raw output: a full three-hour video transcribed line by line, exactly as it comes off the device.
Demo video
See it in action.
A short walkthrough of the device transcribing and answering questions about a live scene.
Agentic demo
Agents at work.
An AI agent using the device's MCP tools to search the footage and answer questions on its own.
Sample transcript
A full 3-hour capture.
The complete, timestamped transcript of a three-hour live video — the raw record you and your agents would search.
For AI agents
Give your agents eyes — and memory.
The device speaks MCP, the open standard for connecting AI agents to tools and data. The same live transcript and semantic search your team uses by hand are exposed as a set of MCP tools, so any agent — in any framework — can watch a site, ask what happened, and act on it. The intelligence stays local; the agent just calls a tool. Use whatever model you want — a local model for full air-gap privacy, or a cloud model like Claude — and it stays cheap: agents query the distilled record, not raw video, so you skip the dominant cost of streaming footage through a model.
search_record()
Find any moment by meaning — “a forklift near the loading door,” “someone in the server room after hours.”
get_transcript()
Pull the plain-language record for any time range, ready to read or reason over.
watch_for()
Register a standing condition in plain English and get notified the instant it happens.
recall()
Ask what happened — “was the gate left open today?” — and get an answer grounded in what was actually seen.
The result is monitoring that reasons instead of just recording — an agent that can connect "the gate alarm" to "the truck that arrived twenty minutes earlier" and tell you what it means, grounded in what the cameras actually saw.
Built for
Industrial & manufacturing
Line safety, PPE compliance, dock and yard activity — monitored continuously without streaming the floor to anyone.
Utilities & critical infrastructure
Substations, pump houses, and remote sites that operate under air-gap or no-cloud policy.
Data centers & secure facilities
Tailgating, access events, and cage activity, kept entirely inside the perimeter.
Healthcare, eldercare & retail
Sensitive environments where the footage simply can't leave the building. Expansion markets as the platform grows.
Privacy & security
Privacy isn't a setting here — it's the architecture. Footage is understood and indexed where it's captured and never uploaded. There's no cloud account to breach, no vendor holding your video, and no recurring stream of your operations leaving the property.
That makes the device a fit for on-prem and air-gapped environments that cloud-based video AI structurally can't serve — and it means your AI agents can have rich situational memory without widening your attack surface.
Cost
Cheaper to run than a lightbulb.
The hardware is a modest one-time cost — but the real savings are in running it. It draws less power than a lightbulb and carries no cloud bill at all, where cloud video AI charges hundreds of dollars a month just to keep processing and indexing your footage.
One-time hardware cost — a capital expense, not a subscription.
Less power than a standard 60-watt bulb. Cheap to run, all day.
No egress fees, no storage bills, no per-camera monthly licenses.
No internet or subscription required to keep it running.
About
OnLookerAI is a deep-tech company building the on-device foundation for private, agent-ready video intelligence. We do the hard systems work — running modern vision-language models entirely at the edge — so that powerful, fully-local video understanding can run on small, affordable hardware.
Founded by an MIT alum with a background in low-level systems and applied machine learning. We're raising a pre-seed round to bring the platform to its first production deployments.
Get in touch
See it running on real hardware.
For buyers: request a demo and we'll show you live transcription and search on your own footage. For investors and partners: read the pitch deck, then reach out.
Or email us directly at babs@wanyekitech.com.