n8n's Assistant asks for a credential at the node that needs it

A credential prompt that arrives at the step that needs it is a different act of consent than a credential prompt at session start ^[2]. The reviewer can see which node is asking and what it is asking for. That ordering is the part of this design worth copying, and it costs nothing to implement beyond deferring the request.

What the gate bounds is timing. Access requires explicit confirmation, as does activation ^[3]. n8n's documentation says credentials are never exposed to the AI ^[4], and handling stays in the standard credential screens rather than in the chat ^[5]. The material describes narrowing when a credential can be used, not what it can do once approved: the reachable surface after approval is whatever that credential already permits in the connected system.

The activation gate rests on a real distinction: a workflow can be harmless as a draft and consequential once it starts reading data, sending messages, updating records, or triggering other connected systems ^[17]. n8n's framing also puts testing and debugging with the assistant on the near side of that gate, with activation left as a deliberate user decision ^[18]. Whether a test run using an approved credential writes to the live system it is connected to is not addressed in the supplied material. n8n says the controls do not remove the need to inspect the workflow ^[19], and that the preview may make errors or change as it is refined ^[14]. Before I approved a credential to a production CRM, I would want the test-execution semantics in writing.

Then the meter. Consumption is token-based against AI credits allocated to the plan, counted separately from the previous AI Workflow Builder ^[7]. n8n notes that provisioning and debugging can consume more credits, so usage depends on how much assistance a workflow needed rather than on whether it was ultimately deployed ^[8]. That is the honest way to describe it and the awkward way to budget it: the cost tracks difficulty, and difficulty is what you were unable to estimate in the first place. The separate pool is good engineering. Two builders drawing on one balance would leave you unable to tell which one drained it.

The arithmetic stops there. Credit allocations for Starter and Pro, top-up prices, and purchase limits are not in the announcement ^[10], so no per-workflow figure can be computed from it. During preview, teams needing more can upgrade or buy top-ups, with further top-up options promised in the coming weeks ^[9] - a familiar unit of time.

Eligibility narrows the audience further. Preview is on n8n Cloud for Starter and Pro ^[11], and on self-hosted Community, Registered Community, and Business from a supported version the material references but does not name ^[12]. Three deployment categories sit outside it: Enterprise, which is on the roadmap, Enterprise Cloud, and self-hosted npm installs ^[16].

The number worth instrumenting during a pilot is credits per workflow that reached activation, held against credits spent on drafts abandoned in debugging. Given how the meter is described ^[8], the second figure is the one that decides whether this pays.