Vaelen: Self-Hosted AI Camera System, No Cloud, Open Source

ambient perception

Something happened on your camera. Now you’re scrubbing through two hours of footage, squinting at grainy clips, trying to remember roughly when it was. Every camera system does this to you.

Vaelen doesn’t. You just ask it. “Was anyone at the door yesterday afternoon who wasn’t a delivery?” Type that into the dashboard, and Vaelen reads everything it’s seen and heard, and answers you in a sentence — with the actual clip linked right underneath, so you can check it yourself in two seconds.

It runs on one machine you already own, with the GPU you already have. No cloud, no subscription, no account, nothing phoning home anywhere. Ask it in English, German, Finnish, or a dozen other languages — it answers in the same one.

Just ask it

This is the whole point of Vaelen. One question box on the dashboard — not a search bar with filters and dropdowns, an actual question, in your own words. Vaelen searches everything it’s recorded by what it means, not by keyword, then has a local AI read the matching footage and answer you directly.

And it’s a conversation, not a one-shot lookup:

• It remembers what you just asked. Follow up with “what color was the car?” and it knows exactly which car. The whole back-and-forth stays visible above the answer, so you can jump back to any point in it.
• It tells you exactly where to look. Not just which recording — the actual minute. Click it, and the video jumps straight there.
• It only shows you what the answer is actually based on. No pile of loosely-related clips to sort through yourself.
• You can stop it anytime and keep whatever it’s found so far — a hard question can take a few seconds to think through, and you don’t have to wait it out if the first answer’s already enough.

If the first answer isn’t sharp enough — four recordings match “car,” say, but none mention the color — just ask again, more precisely. Vaelen goes back to the actual video frames with your new question instead of recycling the same generic description.

Zoom in like a detective

Here’s a problem every AI camera system quietly has: a two-hour recording only gets analyzed at roughly one frame every two minutes. Fine for a general summary. Useless for catching something quick.

High-Density Analytics fixes that. Mark out a piece of any recording — ten minutes, five, thirty seconds — and Vaelen re-analyzes just that stretch, at full density. A one-minute clip gets checked almost frame by frame.

And you can keep narrowing:

Every recording → the 4 that actually match → the 1 that’s right → a 5-minute slice of it → a 1-minute slice, zoomed all the way in.

Each step hands only what matched straight to the next one. Found the moment you wanted? Save it to the recording’s description with one click, so a normal search turns it up from then on — and it even marks the stretch you’ve already zoomed into, so you never redo the same work twice.

Find it, bookmark it, see it before you even hit play

Understanding a recording is only half the job. Getting back to the exact second that mattered is the other half.

• Star it. One click favorites any recording — or a whole archive folder — and a dedicated Favorites tab pulls everything you’ve starred into one place, sorted however you like.
• Drop a marker. Bookmark the exact second something happens, with your own note — “visitor rings twice,” “car reverses into the gate” — then jump straight back, or send someone else directly to that one second instead of making them scrub the whole clip.
• See where the sound is before you press play. A volume curve sits right under every recording — click anywhere on it to jump there. Turn up the volume and the curve grows with it, so a quiet sound that was almost invisible becomes easy to spot, without the loud parts drowning everything else out.
• Fix bad audio without touching the file. A camera with audio that’s slightly out of sync, a mic that’s too quiet, background noise that needs filtering — all fixable per recording, purely on playback. The original file never gets re-encoded.

It doesn’t just watch. It understands.

Everything Vaelen notices comes together into one connected read on what happened — not four separate facts that happen to share a timestamp:

• Sees: people, vehicles, packages — the main trigger for recording at all.
• Hears: a sound model that reacts to whatever you type — “glass breaking,” “dog barking” — no training required.
• Listens: live speech transcription, so one spoken word can trigger a recording by itself.
• Notices: a fall, raised hands, loitering, sudden movement, someone standing too close, where someone’s looking — six signals from one pass, at no extra cost.
• Reads: house numbers, package labels, license plates — matched to people and vehicles it keeps recognizing forever, even after the original video is gone.

All of that feeds into one description: “a person arrives at the door, the microphone picks up glass breaking moments later, and a package on the step reads ‘DHL Express.'” Not four logs. One account of what happened.

And you can brief it like you’d brief a person. Tell Vaelen once who it’s actually watching for — a security guard, a hotel manager, a worried parent — and every judgment call it makes reflects that from then on. Same cameras, same footage. What changes is what it decides is worth mentioning.

Nothing leaves your network. Ever.

Detection, the vision model writing descriptions, transcription, face recognition, the conversational search — all of it runs locally, on your own GPU. A recording never gets uploaded anywhere. A question you ask about it never leaves your house. One self-hosted box does the whole job — open source, no subscription, no account, ever.

How it stacks up

	Motion / MotionEye	Frigate NVR	Vaelen
Core idea	Record on pixel change	Record on object detection	Understand what happened, then let you ask about it
Ask it a real question	No	Semantic search	Semantic search + an AI-written, sourced answer, in whichever language you ask
Remembers the conversation	—	—	Yes — a dozen rounds of follow-up stay in context
Zoom into a specific moment	—	—	Re-analyze just a segment, down to minutes or seconds
Bookmark a moment, see the sound before you play it	No	No	Markers with your own notes, a clickable volume curve under every clip
Starred favorites across the library	No	No	One click, recordings or whole folders, own tab
What triggers a recording	Pixel difference only	Object detection only	Object detection + audio + motion + spoken words + falls/distress
Hears what it can’t see	No	No	A sound model trained on whatever category you type, zero training data needed
Reads text on screen	No	No	Any visible text — signs, packages, plates — not just license plates
Face recognition	No	Third-party add-on	Built in, identity survives even after the video is deleted
Vehicle / plate recognition	No	No	Built in, matched to named vehicles the same way as faces
Anomaly detection	No	No	Learns each camera’s normal pattern automatically
Configurable AI perspective	No	No	Brief it like a person — security guard, hotel manager, worried parent
Runs locally, no cloud	Yes	Yes	Yes — plus a local, conversational AI on top
Cost	Free	Free	Free, self-hosted, no subscription

Plus, for the more technical among you: a native MCP server to run an AI agent against your footage, shared GPU coordination across multiple Ollama endpoints, Immich-compatible XMP export, native MQTT auto-discovery for Home Assistant, multi-account permissions, and a watchfolder/API for batch-processing footage you already have.

Choose Motion for a lightweight daemon with no GPU involved. Choose Frigate for a mature, broadly-supported NVR with wide hardware support. Choose Vaelen if what you actually want is an answer — not a pile of footage to review yourself.

Frequently asked questions

What is Vaelen?
A free, self-hosted, AI-powered camera system. It records on detection and actually understands what it recorded, instead of leaving you to scrub through raw footage to find anything.

Is it really free and open source?
Yes. Full source on GitHub, self-hosted on your own hardware. No subscription, no monthly fee, no account.

Does it need the cloud or an internet connection?
No. Detection, analysis, transcription, face recognition, and the conversational search all run locally, on your own GPU. Nothing leaves your network.

Can I actually ask it questions about what my cameras have seen?
Yes — that’s the core of the product. Ask anything in plain language and it answers directly, with the source recording linked, in whichever language you asked in. Follow-ups stay in context, and you can zoom into a single minute of a two-hour recording if that’s what it takes.

What hardware do I need?
Any NVIDIA GPU from Turing (RTX 20-series) through Blackwell (RTX 50-series) works unmodified. A GPU makes the AI analysis fast; recording itself works even without one.

How is it different from Frigate or Motion?
Frigate and Motion are recording systems with detection bolted on — you still watch and search the video yourself. Vaelen treats the footage as raw material for a small team of AI models working together, so you get a direct, sourced answer to a plain-language question instead of a list of clips to check by hand.

Get it

Bare-metal or Docker install instructions, the full settings reference, and the complete source: Vaelen on GitHub.

---

Ambient perception was supposed to be the modest part. It’s not supposed to notice this much — it does anyway.