Ask mode
Read-only chat: explain a workflow branch by branch, find why a run failed, or quote a step's output. Its tools are filtered server-side; it cannot edit.
Ask is the chat with the keys taken away: it can read everything and change nothing. Use it when you want answers, not edits, especially on workflows you didn't build.
Turning it on
The mode pill in the chat (bottom of the command bar, or the pills row of the docked transcript) is a two-item dropdown: Build, "plans, builds and verifies, or edits what's there", and Ask, "answers only, never changes anything". Pick Ask and the pill's dot turns blue.
Ask is a per-session overlay, not a property of the workflow. It rides on each message you send, so flipping the pill mid-conversation takes effect on your next message, and nothing is written to the workflow. Pick Build again and you land exactly where the build left off: the plan, the verified steps, the pending edits are all untouched. The AI never switches the mode itself.
What it's good at
- "What does this workflow do?" gets a plain-language readout of the graph, branch by branch, from the live canvas the AI sees in every turn.
- "Why did last night's run fail?" makes it read the five most recent runs with every step's input, output and error, so it can point at the step where the data dropped.
- "What did the extract step output?" returns actual run data, quoted.
- "Where is this build up to?" works too: the build plan stays visible to Ask, with each step's status.
- "What would I need to change to…" gets advice, without touching anything; flip to Build when you want it done.
The guarantee
Every tool the AI can call carries a mutates flag. When a message
arrives with Ask on, the server builds the tool list with every
mutating tool left out before the model sees it. What remains on
the canvas surface is reads: the workflow list, the conversation
history, recent runs with step detail, step schemas and output
paths, the step catalog and operation search, the workflow
validator, runbooks, webhook payload research, a URL fetch and a
web search. Staging a graph, editing steps, connecting providers,
running steps, activating, and the plan proposal are all gone.
The rules the model reads say the same thing: answer from the canvas, the plan, the runs and the catalog; when the honest answer is a change, describe it and say that flipping the pill to Build would do it; never present a described change as done. Because the filter is server-side, no phrasing of a request can route around it. That makes Ask safe to hand to teammates who should explore without breaking things.
What Ask cannot do
- It cannot run a step or the workflow to find something out; it reads what past runs recorded.
- It cannot validate-and-fix. It can tell you the validator's findings, not apply them.
- A plan submit (the goal box in Build) never rides an Ask turn.
FAQ
Can Ask see other workflows? Yes, within your workspace. It's the same visibility you have.
Why does Ask refuse to "just fix it"? By design. Fixing is Build's job, and the mode switch is always a human action.
Does Ask cost credits? Yes. A turn is a model call like any other and meters to the workspace the same way.
Is there an Ask mode over MCP? No. MCP has no modes, only token scope: a read-scoped token is the external equivalent, see MCP.