What a team gets
Everything in the personal app, shared.
Shared workspaces with roles
Owner, admin, editor, commenter, viewer. A company or family space sits next to each person's private system, and personal pillars never leave the machine they were written on.
Live co-editing
Shared notes edit together in real time, with cursors and presence, and every page keeps a history of who changed what and when.
Alignment, briefings, and audits
See how every project ladders up to an area and a pillar, who owns it, and where the gaps are. Executive briefings, client status, risk registers and project audits, written from the team's own work.
Assignments and decisions on the record
Assign a task to a teammate, request a decision on a page, and get told in the app. Decisions are filed under the project they belong to, not lost in a thread.
An admin console
Members, an access matrix across every space and person, and the audit trail. For Enterprise, single sign-on enforced at sign-in and SCIM provisioning built against Microsoft Entra. Security has the detail.
Agents, with a person in the loop
Connect any agent over MCP. Its work arrives as proposals you approve, narrow, or deny, and every decision is kept, attributed, and exportable. The whole loop is below.
The alignment view · no project without an owner or an objective · captured from a demo workspace
A project page · owner, stage, deadline, and decisions filed where the work lives · captured from a demo workspace
Everyone also keeps the personal app: their own vault, offline, free. What one person gets →
Agent work
Your agents propose. You decide.
When you hand work to an agent - yours, ours, or any agent that speaks MCP - it does not write into the workspace. It proposes. A person with the right role approves it, narrows it, or denies it, and the decision is recorded with who, when, and why. The rest of this page is that loop, up close.
Connect
One endpoint, one key. Your agent keeps the tools and the framework it already has. The key resolves to exactly one agent in exactly one workspace, and nothing in a request can name a different one.
Remote MCP · Live
Land
Its writes become proposals instead of changes. Each one carries what it wants to do, the evidence behind it, and the confidence it reported. Nothing has moved in your workspace yet.
Proposal queue · Live
Decide
Approve it, untick the parts you do not want, or deny it and remember that answer as a standing rule. Your verdict is the write.
Web and desktop · Live
Grounded, Not Guessing
It reads the workspace before it answers.
One real turn: asked what is at risk this week, the assistant reads the alignment view, lists tasks across every project, and checks the linked issue tracker - then names the risk, the owner, and the decision that is blocking. And when a connection needs re-authorizing, it says so instead of inventing an answer.
Eight steps, four tool calls · captured from a demo workspace
Nothing lands until a decision is recorded.
The proposal
The proposal, up close.
confidence 0.82
Draft the v0.1.42 release note and file three follow-ups
- Create pageRelease notes / v0.1.42
- Add taskChase the Windows signing ticketdue 2026-08-27
- Add taskRe-run the updater canarydue 2026-08-22
- Update pageChangelogbody replaced
You can narrow it. Untick a line and only what is left runs. An approval can remove items. It can never add one, by construction, and the server refuses a widened approval outright.
You can say why. A verdict carries the reason you typed alongside who gave it and when, so the record answers why as well as what.
A denial can stick. Remember an answer for a set number of days and matching proposals are refused at submission. The desktop app and the web queue both list your standing rules and let you or an admin remove one.
Nothing is decided twice. Two admins opening the same proposal is a race the first verdict wins. The second is told which verdict stands rather than silently overwriting it.
You can let the safe ones through. Auto-approve is off until you switch it on. Turn it on with a confidence floor and a batch that clears the floor applies straight away, recorded as approved by your policy rather than by a person, with the reason on the row.
The record
What happened, and who said yes.
A per-session approval prompt is gone the moment the session closes. A verdict here is a row: attributed, timestamped, and readable by the people accountable for the work.
Every proposal keeps what was asked for and what was approved, side by side, so a narrowed approval never overwrites the original request.
Every verdict records who decided, when, and the reason they gave.
A digest of the batch is taken at submission and again at the verdict, so you can tell that the work a person decided on is the work the agent sent.
Applied and approved are recorded as different facts. A proposal that applied three of five changes says so.
Agent actions in a shared space are readable by that space's admins, not only on the machine that ran them.
An agent key is capped at viewer, commenter, or editor. Managing members and deleting a workspace are not capabilities an agent can hold.
Aurum holds no compliance certification today. The above is what the product records, not an attestation. See Security.
The async contract
Plan for hours, not milliseconds.
A verdict arrives when a human acts on it. That might be four minutes or it might be the next working day, and the loop is built for the second case. Your agent submits, gets a queued acknowledgement, and goes away. It polls later, from a fresh process if it likes, and picks up every verdict it has not seen.
Losing a connection, a process, or a whole day loses no work. A proposal nobody answers expires on a stated date rather than sitting open forever, and the expiry is itself a recorded verdict.
SUBMIT → queued, immediately
POLL → every verdict since your cursor
NO ANSWER → expires on a stated date, recorded