Tanuki

Tanuki turns the messy inputs a small business generates all day — voice notes, meeting recordings, PDFs, half-written drafts — into clean, structured text. It runs entirely on your own hardware. Nothing is uploaded, nothing is metered, and there is no per-seat cloud bill that grows with how much you use it.

It comes in two shapes, and they are genuinely different products for different situations.

One: the desktop app

A normal Mac app from the App Store, for one person on one machine.

Drop a file on the workspace and Tanuki filters its workflow list to the ones that suit that input. Drop a voice recording, pick Transcribe and Tidy. Drop a contract, pick Summarise Document. Press run. The output lands in the app, and optionally gets pushed straight to wherever you keep things — S3 or MinIO, WebDAV, Dropbox, iCloud Drive, Google Drive, OneDrive, or a folder on your Mac.

There is a Compose view for working live rather than in batches: dictate an email, paste in a rough paragraph, and run a workflow over it to fix the language and structure it. It reads instructions embedded in your own text, so ending a brain-dump with “clean this up and make bullet lists where appropriate” is a valid way to drive it.

It also keeps chat history locally, indefinitely, because text is cheap on a hard drive. That makes “last year we discussed suppliers, which ones came up?” a question you can actually ask — a thing cloud assistants with rolling context windows can’t do.

Two: the network appliance

The same engine with the interface removed, running as a background service on one Mac that the whole office talks to over HTTPS.

This is the shape that suits a small business best, and it is deliberately unglamorous: a Mac Mini in a cupboard. It starts at boot, has no window, no Dock icon, no menu bar, and needs no one logged in. Staff reach it through a small web front end, or through whatever you already use — it exposes a documented HTTP API with a Swagger interface for testing.

The appliance is not a scaled-down cloud deployment. It is a single machine, by design, and that constraint is what makes it cheap and predictable.

What it actually costs to run

We load-tested the appliance hard before we would say any of this, and published the numbers in full.

32 to 37 watts under continuous heavy inference. That is roughly $30–50 a year in electricity at commercial rates, and that assumes it is hammered around the clock — a real office with idle periods costs less. Apple silicon draws essentially nothing at idle, so a machine waiting for work is not costing you anything.

Twelve hours unattended, 1,742,332 requests, zero failures. Memory settled at 8 GB after two hours and drifted under 200 MB across the following nine and a half. It does not need rebooting, babysitting, or a person who knows what a GPU cluster is.

It stays quick when several people use it at once. Four staff running full audio workflows continuously for fifteen minutes, with status polling throughout: worst single response across the entire window was 26 milliseconds. It doesn’t get weird on a busy day.

A closet is fine. 18–23 °C of thermal headroom under the heaviest load we could produce. Fans only become audible after about nine minutes of sustained work. No special cooling, no noise complaints.

The whole cost model is different from a cloud subscription: the hardware and the power are the cost. A workflow you run once costs the same as one you run ten thousand times. Capacity planning is arithmetic — a 7B model needs about 8 GB of unified memory all-in, which fits a base 16 GB Mac with room to spare.

Why it stays on your premises

Every file stays on the machine you own. No tokens leave the device, and after the initial model download nothing in the speech or inference path needs the internet at all — it will run air-gapped.

For a small business that mostly means the awkward conversations stop. Client audio, HR matters, medical or legal records, anything under contract: the compliance question shrinks from “where does this data go, who can see it, and what does their sub-processor list look like” to “it’s on the box in the cupboard.”

Workflows handling genuinely sensitive material can be marked ephemeral — they process, deliver the output to wherever you configured, then delete the input, the intermediate steps and the output from disk. Nothing to find later, because nothing was kept.

What it is not

It is not a frontier model. The default is a 7B open-weight model, and it does what a 7B model does.

That is a choice, not a limitation we’re apologising for. For the work small businesses actually need — cleaning up dictation, structuring notes, summarising documents — smaller models do a better job. They stay closer to what was actually said instead of embellishing it. We’re not asking it to be creative; we’re asking it to be accurate and fast.

If you need frontier-model reasoning, keep using a frontier model. Tanuki handles the constant, high-volume, low-drama text work that would otherwise cost you a subscription per head — and it complements that rather than replacing it.

Why we can say any of this

We started in January 2026 with no Apple experience. The desktop app shipped on the 1st of July, we began the mobile version on the 2nd, and it passed Beta Review on the 5th.

That was possible because the suite is built on hexagonal architecture: the core knows nothing about how it is being driven, so a new platform is a new adapter rather than a new project. The desktop app and the headless appliance are the same inference engine, the same HTTP API and the same workflow runtime behind two different front doors.

CyberBench is the third branch of the same platform, built for corporate security teams and sold direct.