Documentation
Intelligence Nucleus
The reasoning layer that sits above all nine products. You state an intent; the Nucleus turns it into work, runs it across domains, checks its own results, remembers what matters, and can act on your behalf -- entirely on your machine.
Why it beats a bigger model
A better assistant is not just a bigger model. The Nucleus wraps the model in structure -- planning, checking, correcting, and remembering -- so the quality comes from the system, not only from raw model size. The practical result: a small model you can run locally, working this way, can outperform a much larger model used on its own.
What it does, end to end
Every request flows through the same shape, whether it is one quick answer or a multi-step job spanning several products:
It reads your intent in context -- what you are working on and what you have asked before.
It decides the work that needs to happen and in what order, across the right domains.
It carries the work out across mail, money, projects, the web, and more -- in parallel where it can.
It checks each result before using it, and corrects course on its own when something falls short.
It returns the result where you asked for it -- on your desktop or in a messaging channel.
Two models, working together
You can run two models at the same time: one to plan and write the final answer, another to do the faster work alongside it. They run in concert, so thinking never waits on doing -- and the second model can check the first before its work is accepted. Both are local models you choose, and you can swap either one whenever you like.
It adapts to the task
Simple questions get a direct answer with no overhead. Bigger jobs get a real plan, run several agents in parallel, or loop and refine until the result is good enough -- and hand off cleanly between products when the goal belongs somewhere else. You never have to pick the mode; it chooses the lightest path that does the job.
It checks its own work
Results are verified before they are used. When something is empty, malformed, off-track, or unsupported, the Nucleus retries and, if needed, re-plans -- so you can trust what it finally gives you. For research, claims have to trace back to a real source to make it into the result.
You decide how much it can do
Autonomy is set per area of work, so email can stay cautious while research runs free. Approvals are durable -- a request waiting for your sign-off survives an app restart, and you can approve it later from your phone.
Every action waits for your confirmation.
It acts on its own for read-only work, and asks before anything with a consequence.
Once it has proven reliable for a kind of action, it runs it automatically -- and asks again whenever it is unsure.
It always confirms first -- reserved for sensitive lanes like sending email.
It remembers and gets sharper
The Nucleus remembers what is true about you and your work, what happened and when, how you like things done, and what is still in flight -- and it surfaces the right context even when you phrase things differently than before. It understands meaning, not just keywords, and all of it stays on your device.
It works while you are away
The Nucleus is not only reactive. It can run recurring work on a schedule -- a morning briefing, a follow-up, an overnight research job -- and send the results to your desktop or a messaging channel. You can also command it from Telegram, iMessage, and more, with the same approvals and safeguards as the desktop app.
ships the Intelligence Nucleus on your own machine. Get notified when the first macOS build goes live.