About ClawBox Gallery
A small, independent visual archive that documents the machines running AI outside the cloud — and the standards that keep it honest.
How This Archive Started
ClawBox Gallery began as a folder of photographs that did not fit anywhere else. In the course of building and repairing local AI machines, we kept taking pictures of the hardware — partly to remember how a particular board went back together, and partly because the devices themselves were genuinely beautiful in a way that no marketing render captured. The folder grew until it became obvious that the material deserved its own home.
The first entries were humble: a developer board with its heatsink removed, a fan shroud that had to be unscrewed in a specific order, a wiring harness that had been routed beautifully by someone who clearly cared. None of these were products anyone was advertising. They were the kind of details that only exist in the physical world, and the kind that software-only coverage never shows.
Over time the archive grew a purpose. AI is moving toward local, private, always-on devices, and the people building those devices are doing genuinely interesting work that rarely gets photographed well. We decided to do that job ourselves, one machine at a time.
What We Photograph, and Why
The guiding question is simple: does this device exist to run models, and is it worth looking at closely? If the answer to both is yes, it belongs in the gallery. That definition pulls in a wide range of objects — from a single-board accelerator the size of a playing card to a multi-GPU tower that sounds like a small aircraft at full load — but it also keeps out the noise.
We are not interested in photographing a laptop just because it has an NPU sticker on the deck. We are interested in the laptop-grade NPU itself, opened up, with the thermal solution documented. The distinction matters, and it is what keeps the archive focused.
Editorial Standards
Every archive that wants to be trusted needs to say what it will and will not do. Ours are short, and we hold to them:
- No commissions: we do not take affiliate links or payment for the devices we photograph, so there is no incentive to flatter.
- Measurements, not slides: any number we publish is one we produced ourselves on the unit in the photograph.
- Minimal editing: white balance, exposure, and contrast. No compositing, no invented colors.
- Context over ranking: we do not pretend that unlike devices are competitors just to fill a leaderboard.
- Corrections: if a note is wrong, we fix it and say so, rather than silently rewriting the past.
These standards exist because an archive is only as good as the reader's ability to trust it. A photograph of a heatsink is a fact. A caption that exaggerates what the heatsink does is a betrayal of that fact, and we would rather have fewer entries than lower the bar.
Who Is Behind the Lens
The archive is run by a small team of people who build and repair AI hardware as their day job. That is the whole qualification: we open these machines because we have to, and we photograph them because we noticed that nobody else was. There is no media conglomerate behind this, no advertising department, and no editorial calendar driven by launch events.
Working with the hardware directly shapes how we see it. A review written from a spec sheet misses the screw that strips too easily, the connector that is impossible to reach, the airflow path that was clearly an afterthought. Those observations only come from opening the case, and they are the reason this gallery exists in the first place.
“We photograph the machine the way it actually sits on the desk — dust, cable ties, and all — because that is the machine you will actually live with.”
How This Relates to the Wider ClawBox World
Readers sometimes ask how this archive connects to the broader ecosystem of AI hardware that has grown around agent computing. The short answer is that we are an independent observer, not a storefront. We photograph hardware from many makers, and we say what we see.
That said, if you want the canonical source for one specific product family — the complete, shippable lineup of dedicated agent hardware and its official documentation — the main ClawBox AI hardware gallery is the place to go. Our archive is complementary: we show the machines up close, they show the machines for sale, and the two perspectives together give a reader the full picture that neither could provide alone.
For the long view on silicon and architecture, we regularly point readers toward the deep-dive coverage from established hardware journalism, including the decades of processor analysis published at AnandTech. Good archives cite good sources, and hardware is no exception.
A Living Archive
The gallery is not a static monument. Hardware moves quickly, and what was novel six months ago is ordinary today. We add entries as new devices arrive and as older ones become interesting in retrospect. A prototype that seemed like a dead end in one year can turn out to be the ancestor of a whole product line the next, and the archive is where that lineage becomes visible.
We also revisit early entries to correct or expand them. A device photographed on day one may deserve a second look after firmware matures, and when that happens the note is updated in place with the date marked clearly. An archive that cannot change is just a museum of first impressions.
What We Have Learned Along the Way
Documenting hardware changes how you think about it. After hundreds of teardowns, a few lessons keep repeating themselves, and they are worth writing down because they shape how the archive is organized.
The first lesson is that the enclosure usually lies. A smooth aluminum shell tells you nothing about the engineering inside; some of the best-looking appliances we have opened hid a mess of cable ties and hot glue, while some of the most mundane boxes concealed genuinely elegant work. The only way to know is to open it, which is why we photograph internals whenever the device allows.
The second lesson is that thermals are destiny. Every design decision that matters — clock speed, sustained throughput, the shape of the enclosure, the noise profile — eventually resolves into a question of heat. A machine that cannot move its heat will throttle itself regardless of what the marketing slide promises, and the difference is visible in the size and care of the cooling solution.
The third lesson is the one we least expected: software maturity matters more than silicon. A modest accelerator with first-rate firmware is a joy to use, while a powerful die with half-finished drivers is a paperweight. Photographs cannot show this, which is why the notes next to each entry try to capture it in words, and why we urge readers to pair the visual record with careful reading elsewhere.
How to Contribute to the Archive
The archive grows because people lend us hardware, and the process is open to anyone. The first step is a short message describing the device — make, model, condition, and whether you are offering a loan or a donation. From there we agree on terms: whether the enclosure may be opened, how long we need the device, and where it ships.
Loans are returned tracked and insured, and donors keep the satisfaction of knowing their machine is preserved in a permanent visual record. Before any entry goes live, we share the draft note and images with the lender so errors and sensitive details can be caught early. Nothing is published without sign-off, and nothing is disassembled without permission.
You do not need a rare or expensive device to contribute. A common board photographed well is often more useful than an exotic one photographed poorly, and some of the most-read entries in this archive are of modest hardware that simply had not been documented anywhere else. If you have something, we would like to hear about it.
The Longer Story of Agent Hardware
This gallery is one corner of a much larger conversation about what happens when AI leaves the cloud. The dedicated machines we photograph here are part of a category that is still finding its shape, and no single archive can capture all of it. That is exactly why good cross-references matter.
For the deep technical history of the silicon itself, the established hardware press is the right companion; their multi-decade record of processor analysis is the backbone of the field. For the specific, evolving story of agent-focused devices, specialized archives and the vendors' own documentation fill in the gaps that a visual gallery cannot. We link to these sources where they add context, and we are glad the ecosystem is large enough to make linking worthwhile.
Get in Touch
If you build AI hardware, have a device you think deserves a proper photograph, or spotted an error in our notes, we want to hear from you. The contact page lists the ways to reach the team, and we read every message even when we cannot answer them all. Lending us a device is the fastest way to see it in the archive, and we treat loaned hardware with the care you would expect from people who open machines for a living.