How Magieva works¶
Why this page exists¶
Magieva behaves differently from a chat assistant, and the differences are deliberate. Understanding the shape of the system makes it obvious why some requests answer instantly, others become tracked work, why the companion sometimes acts before you ask, and why the organization boundary is drawn where it is.
A companion inside an organization, not a chatbot beside one¶
Most AI tools sit outside your business and answer questions about the world. Maggie sits inside one organization and answers questions about that organization.
Everything she can see — your documents, your projects, your connected tools, your people and their titles — belongs to that organization. This is what lets her answer "what does our refund policy say?" rather than "here is what refund policies usually say".
It is also why an empty organization produces disappointing answers. There is nothing wrong with the model; there is simply nothing of yours for it to reason about.
Requests become either replies or work¶
When you ask for something, Magieva first decides what kind of request it is.
- A question gets answered in the conversation.
- A piece of work becomes a task.
That decision is why the product has a Tasks view at all. Work that takes minutes should not be trapped inside a chat window where closing the tab loses it. A task has its own record, its own phases, and its own log, and it survives you walking away.
The practical consequence: if you asked for something substantial and the conversation looks quiet, the work is almost certainly running in Tasks.
Work runs as a sequence of steps you can watch¶
A task is not a single call to a model. It moves through phases — gathering context, doing the work, checking it, producing output — and each transition is recorded.
This is why you can open a running task and see where it is rather than a spinner, why a task can pause to ask you a question and then resume, and why tasks can depend on one another. It also means a task that fails does so at an identifiable point, rather than simply not working.
She can go first — inside a boundary you set¶
New in release 26.37
Live since September 10, 2026. Proactive behaviour is switched on progressively; reaching you off the app happens only with your consent.
From release 26.37 Maggie does not only respond. She can do a useful first thing before you have asked, keep a watch on something for you, and propose actions of her own. The design principle is that initiative is bounded, and the boundary is yours: unprompted work draws on a separate, lower spend ceiling; anything she proposes on her own either matches a standing rule you set or waits for your answer; irreversible actions and anything that moves money always ask; and every unprompted action can be traced back to the signal, the rule and the cost that produced it. She reaches you off the app — by text or call — only after you have said she may. The full picture is in A companion that goes first.
Everything is scoped to one organization¶
The organization boundary is the strongest line in the product. Conversations, projects, tasks, files, integrations, and your token balance all live inside exactly one organization and do not cross.
This is a data-protection property first: two organizations using Magieva cannot see each other, even indirectly through Maggie's memory. It is also why moving work between organizations means recreating it, and why belonging to several means switching explicitly between separate worlds.
Inside an organization, access follows people and groups: a project shared with the Design group is visible to whoever is in Design today, and the organization's activity page shows each person only the entries about things they can already see.
Shared knowledge flows one way¶
There are things every organization needs that are not specific to any of them — the catalogue of available AI models, the directory of connectable tools, published templates. Magieva curates these centrally and publishes them out to every organization.
That flow is deliberately one-directional. Your organization reads from the shared catalogue; nothing your organization does is written back into it. Your material stays yours.
Memory is built from what you give it¶
Maggie's knowledge of your business is assembled from what you connect and upload — documents, repositories, connected tools. Magieva reads that material, breaks it into retrievable pieces, and draws on the relevant ones when answering.
Two things follow. First, a citation is meaningful: she is pointing at the actual source she used, and you can check it. Second, the quality of her answers tracks the quality of what you have given her far more than it tracks anything else.
From release 26.37 all of that material lives in one place, the Library, with three lenses — your files, your knowledge bases, and what she remembers — and one search across them. The Library also tells you honestly how much of it she can currently see by description or transcript, so a quiet miss is never mistaken for "nothing matched". See Find and organise your files in the Library.
Why answers cite their sources¶
Where Maggie used a specific source, she says so and links to it.
This is not decoration. An AI answer you cannot verify is an answer you cannot safely act on. Citations make the difference between output you can forward to a client and output you have to check from scratch — and an answer that should have a citation but does not is a useful warning that she is reasoning from general knowledge rather than your material.
Where to go next¶
- A companion that goes first
- How Maggie chooses a model
- Skills and Tools
- Getting started