The in-product Flowgen Agent turns a plain description into wired, configured, and (with approval) running nodes on the canvas. Open it from Assistant in the left sidebar. Choose Assist for focused work or Plan & Build for a staged flow; the bottom Make composer is for one direct node or output.
Where: Flowgen canvas → left sidebar → Assistant — Click Assistant in the left sidebar, then describe what you want
Key ideas
- Two modes: Assist works on what is already on the canvas and adds to it. Plan & Build proposes a whole flow, shows you what it would create before anything lands, and then builds it in stages so you are never handed a finished machine you did not watch appear.
- Context and references: From the Assistant conversation, add a node already on the canvas, a file from Assets, a fresh upload, a link, or typed text — and label each one as a Reference, a First frame or a Last frame where supported. The one-action Make composer is a separate bottom surface.
- JSON action plan: The model returns a JSON plan of actions (create / setModel / setProps / setPrompt / connect / run). The panel executes them against the canvas API: create the node, apply the model, set props, set the prompt, then run.
- Google-first models: The assistant defaults to the cheapest Google model that fits — gemini-3.1-flash-image for images, veo-3.1-lite for video — and only escalates (e.g. to veo-3.1 when a reference/last frame is present or 4K is asked).
Steps
- Click Assistant in the Flowgen left rail to open the in-product AI Agent. MCP has no tool that opens this panel; an external MCP client uses its own model and direct Flowgen tools.
- Pick a mode — Assist to build on what is there, Plan & Build to design a whole flow first — then type what you want.
- Reply with the concrete description plus any options (duration, aspect ratio, audio) — it sets those via the matching props.
- Watch the assistant create the node, choose a Google model, set the prompt, and run it; a 'Generating…' indicator shows while it works.
- Keep chatting to extend the same flow (the last created node is carried as context), or click the refresh icon to start fresh.
FAQ
Where are the creation tiles the assistant used to open with?
They are gone. Choosing what to make is the Make bar's job now — pick an action there and it builds the node directly, without a conversation. The assistant is for the cases the bar cannot express in one action.
Which model will it pick?
It prefers Google and the cheapest model that fits — gemini-3.1-flash-image for images (gemini-3-pro-image only for best quality/4K) and veo-3.1-lite for video, escalating to veo-3.1 only when a reference or last frame is attached or 4K is requested.