SuperDuck is a superintelligence that belongs to you and nobody else. It sits at a desk in your office and works through whatever you hand it, whether that's a manuscript or a year of invoices. It remembers what you told it last year, and more powerful hardware lets it run larger or faster models.
People have invested in their personal right to approach God without a priest, to read without a censor, to publish without a publisher, and to compute without an operator.
In the mainframe era, a computer was a machine down the hall, or in another building, that you weren't allowed near. You handed your work through a window and somebody else decided when and whether your program ran.
The Personal Computer replaced the mainframe because it was yours, a machine on your desk: you chose what ran on it, nobody counted your keystrokes, and your work filled it over the years.
Now with super intelligence, it's the same script once more. The large AI services are the machine down the hall again and the operators are back. You reach the machine through their window, under their rules, and they keep the transcript.
SuperDuck puts the personal in superintelligence, the way the PC put the personal in computer.
Not because running a model on your own computer is inherently better or cheaper, although it can be, but because nobody should stand between you and your own thoughts. And when you need more compute, you buy more hardware, or you pay a rush fee for an accelerator, instead of subscribing to a service that monitors your thoughts.
We take the word "personal" seriously.
It runs on a machine you bought. You pick the model and you decide when to update it. With local processing, no outside inference provider receives your chats. Optional services are your decision.
In local mode your drafts, notes, memory and ledger live in a folder you choose, in ordinary readable files. You decide how to back them up; synced folders have their own rules.
It works for you, not for a platform. There are no ads and no training on your data, and nobody is tuning it to keep you on longer than you meant to be.
When you're out of the office, your phone connects to your own Mac over an encrypted relay, so you're talking to the same intelligence with the same memory and the same files.
More memory and computing power can let you run larger models or finish work faster. Start with a suitable Apple silicon Mac; larger configurations and team deployments have different requirements.
Your Mac does the thinking, and your phone is just how you keep in touch with it when you're away from the desk.
A chat window hands you a paragraph and leaves the rest to you. SuperDuck can read a whole folder, work through it for as long as the job takes, and leave the finished draft where you'll find it, because it has your files, a memory that lasts, and the time to use them.
refine will work through a whole book, pass after pass, while you get on with something else, and still remember chapter one when it gets to chapter nine.
You put a manuscript on the desk, it reads it, marks it up, writes the next draft and leaves it there for you. You don't paste things into a box and copy them back out.
Tell it something in March and it'll know it in October. Say remember and it writes it to a file you can open yourself.
It reads and writes your files, keeps a ledger, does arithmetic properly, draws on a screen behind the chat, makes images, talks in a voice you picked, and reads a page you hold up to the camera. It comes in a handful of themes too, which matters more than you'd think for something you'll use all day.
It answers a local API in OpenAI format, so VS Code, JetBrains, Obsidian and anything else built to talk to OpenAI can point at SuperDuck instead, and the code never leaves your Mac.
When one Mac is not enough, ask about larger local or team configurations. The point is to choose where the work runs, not to surrender control of it.
Some work involves confidential information entrusted to you. SuperDuck can process that material locally, without sending it to an outside inference provider, when external services are off. Check synced storage, backups, device security and professional obligations before using sensitive material.
Work through a discovery folder locally and review draft briefs, subject to the firm's security, confidentiality and verification requirements.
Work with permitted local material; clinical accuracy, patient privacy, device security and the practice's compliance requirements still govern use.
Your raw data, your drafts and the papers you're citing, worked through from analysis to submission on your own hardware.
Develop an idea all the way to a filing before anyone outside the room has seen a word of it.
A trading strategy you've typed into a rented chat window has already been shared with someone. Here it stays on your desk.
Decks, diligence notes, and a memory of every deal you've looked at going back years.
Board papers, strategy drafts and the quiet conversations, with an assistant that reports to you and not to a vendor.
One intelligence for the whole firm, running on hardware in your own back room that you own outright.
SuperDuck thinks on your Mac. A model you download once, in Settings, works with your folders, your mail, your calendar and your reminders, without sending them to an outside AI provider. For most of the day, that is enough.
When it isn't, Toasted is a switch. Flip it and the thinking moves to a model of your choosing, including frontier models, pay as you go, no seat, no tier. Before a word leaves the Mac, Veil can take out names, contacts, numbers and places, replacing them when the answer returns. The big model does the work; it need not know whose work it was.
Off is the default.
Mail, calendar and reminders may sync through their own account providers. Veil reduces identifiable details before outside processing; surrounding context can still reveal them.
Right now that means Apple silicon, either one Mac or several on your network. The model downloads once and runs there with nothing in between you and it. New models come out all the time, and when one's worth having you update from Settings on your own schedule. With local processing and optional external services off, inference stays on the Mac. Synced folders, backups and network services have their own data paths.
Pair an iPhone or iPad and ask it things from the train. The question goes over an encrypted relay to your own Mac, which does the work and sends the answer back.
In Tools, Internet search is off until you enable it. Toasted can handle a long refine or a model too large for your Mac; with your own OpenAI, Anthropic or compatible API key, you can choose another provider instead. Use a local model for one job and an outside model for another. Your workbench, memory and saved transcripts remain on your Mac, while requests sent for outside processing are handled by the provider you select.
The desk is an ordinary folder on your Mac that you picked. Memory, ledger, notes and drafts are all plain text, so you can open them in any editor and back them up with Time Machine or whatever you already use. If you ever leave, you take them with you.
Free while we're in beta.
When you don't want to wait or need more compute. Toasted offers faster inference and a subscription free, pay-as-you-go gateway to other models, including Claude, GPT, Gemini, and others, subject to their terms of service. Mixing and matching models to your task can be very token efficient.
One superintelligence for the whole team, running on your premises, that nobody outside the firm shares, meters, or trains on.
Some people pay for several AI subscriptions. SuperDuck offers a local workbench at $99/month after beta, while optional Toasted and BYOK model usage is billed separately. The services below do not offer identical capabilities, so this is a comparison of advertised subscription fees—not a promise of total savings. Tick what you pay for now to compare fees.
$0 /month
Approximate public list prices, subject to change; check each provider before relying on them. These are not interchangeable products. SuperDuck software costs $99/month after beta ($990/year if paid annually). Local model use does not incur a third-party per-token fee, but optional Toasted/BYOK inference can add charges. This calculator compares recurring subscription fees only, not feature equivalence or total AI expenditure.
SuperDuck is in open beta on TestFlight for Mac and iPhone. Download it, point it at a folder, and by tonight you'll have an intelligence that's yours.