The assistant
The workspace assistant — opening it, what it can read, which actions wait for your confirmation, how roles apply, flow proposals, and what happens without the model.
The assistant is a chat panel for the whole workspace. Ask what needs attention, find orders, check stock, read a flow, or have it decide an approval, deliver an event, cancel an order, publish a flow or run the stock review. It answers from live data through tools, and anything that changes something stops for your confirmation first.
One assistant serves every page.
Opening it
Section titled “Opening it”- The Assistant button in the top bar, or ⌘J (Ctrl+J) anywhere, which also closes it. On wide screens the panel reopens on your next visit if you left it open.
- The Ask button on an order page or in the flow editor opens the panel about that order or flow.
Enter sends, Shift+Enter starts a new line. Stop ends a reply in progress. A message written while a reply is still streaming is held, shown above the prompt where you can edit or remove it, and sent when the reply ends.
The context chip
Section titled “The context chip”The chip in the panel’s header says what “this” means: the order or flow open on the page, or Everything when you are elsewhere. Open the chip to switch between the page and the whole workspace. The page’s context travels with each message you send. In the flow editor it includes the graph on screen, unsaved changes included.
The attention line
Section titled “The attention line”The top of the panel always shows what is waiting on people, refreshed while the panel is open and after each reply, with links to each item:
- open approvals, oldest first;
- orders that failed or rolled back in the last 3 days;
- orders that have waited on a business event for over a day.
What it can read
Section titled “What it can read”Every role can read. The tools are needs_attention, find_orders (by status, model, channel, customer,
text, total, age), get_order (status, what it waits on, steps run, recent activity),
assess_order (typed questions about one order, answered with a probability for every option; up to 4 images),
find_products (stock, reserved, available, reorder points), list_flows and get_flow (a saved
draft with its validation problems). Text, customer and SKU searches match % and _ literally, not as LIKE wildcards. Reads run without asking, and each shows as a row in the
conversation.
Actions that ask first
Section titled “Actions that ask first”These tools never run on the model’s say-so. The panel shows a confirm card for each; nothing is sent until you press the confirm button, and declining leaves everything as it was.
| Action | Needs | Does |
|---|---|---|
decide_approval |
approver |
Approves or rejects the approval an order waits on, with an optional note |
deliver_event |
operator |
Delivers a business event to the order waiting for it |
cancel_order |
operator |
Cancels the order and undoes what it did, as Cancelling |
publish_flow |
operator |
Publishes the flow’s saved draft as the next version |
run_stock_review |
operator |
Starts the stock review now |
The card is not the model’s summary. It is resolved against live data when it appears, and states the consequence: for a decision, the order, its total and which step it goes on to; for a cancel, the effects it will undo and the reason; for a publish, what changed since the live version and that unsaved editor changes are not included; for the stock review, which SKUs are at or below their reorder point. When the action can’t go ahead (already decided, nothing waiting for that event, the order has finished, the draft has errors, nothing to publish, or your role is too low) the card says why and offers no confirm.
Confirming runs the action as you, after the engine re-checks your role and the order or flow
again. The state can have changed since the card appeared, in which case the action is refused with
the reason. An approval decided here has no 5-second undo window, because the confirm is the
check. The audit trail names the actor as <email> via assistant: that string is the approval’s
by, or ends the cancel reason.
The assistant works within your role (viewer, approver, operator, in order). A tool your role can’t use is
not offered to the model, and is refused if it is called anyway; the assistant says which role can.
Every role can use the read tools. The role is re-read from your session when a tool runs, so a
role change, a sign-out or a disabled account applies to an open panel at once. See
Team and roles.
Flow changes
Section titled “Flow changes”Ask for a change to a flow (“add a fraud check after credit”). The propose_flow_change tool
(operator) drafts a whole new graph with the same copilot that creates flows, validates it, and
returns the explanation, a count of what it adds, removes, edits and rewires, and any problems. It
never saves anything. Review in editor opens the flow’s editor with the proposal, where
accepting it is one undoable edit that you can still save and publish. In the editor the draft
builds on the graph on screen. A request that changes nothing says so, with no card.
Creating a flow from a description still happens on the Flows page.
Without the model
Section titled “Without the model”The assistant answers with @cf/moonshotai/kimi-k2.6 by default (see
AI nodes and the assistant). It falls back to a deterministic rule-based model,
which calls the same tools through the same confirm cards, when AI is unavailable, the model
errors, or nothing useful arrives within 20 seconds.
A reply from the fallback is labelled Answered without the language model, from fixed rules. It understands plain requests (“what needs attention”, “add a risk score after credit check”), not open questions.
Over MCP
Section titled “Over MCP”The same tools are served to MCP clients such as Claude, Claude Code and Cursor at /mcp, with the
same roles. There is no confirm card: the client asks, using each tool’s annotations, and the
engine still re-checks the role and the state before it acts. Writes are recorded as
<email> via MCP. See MCP server. Other agents reach the same tools over the
Agent2Agent protocol at /a2a, as skills; see A2A endpoint.
Privacy
Section titled “Privacy”Each person has one assistant and one conversation, tied to their signed-in account, so nobody can open another person’s thread. The conversation keeps the last 200 messages. Conversation options in the panel clears it. A reply that is in progress when the tab closes keeps streaming and is there when you come back. Everyone signed in with the shared workspace password shares one assistant. Orders, flows and approvals are read from the same data the dashboard shows, so the assistant tells you nothing your role couldn’t see.