Open Assistant from Flowgen’s left rail for conversational help, direct graph building, or an editable staged plan executed with the canonical Flowgen tools. The bottom Make composer handles one direct node or output.
Where: Flowgen left rail → Assistant
Key ideas
- Assist: Use conversation for explanations, diagnosis, and focused requests while retaining session history, model, and thinking controls.
- Plan & Build: For multi-step work, the agent can stage an editable plan, wait for review, then create/connect/configure/run the graph with Flowgen tools.
- Host-owned reasoning: When Flowgen is controlled through MCP, the MCP host model should call Flowgen tools directly; it should not proxy its reasoning through the in-app assistant.
Steps
- Click Assistant in the Flowgen left rail and choose Assist or Plan & Build.
- Describe the outcome, constraints, existing nodes to reuse, and what must remain unchanged.
- For a staged plan, edit/reorder the proposed work before approving the build.
- Watch Activity and the canvas while tools run; inspect connections and outputs before continuing the conversation.
- For a request that only asks to arrange existing cards, the Agent may acknowledge the layout after the arrange action succeeds and it reads back the current graph. That completion applies only to layout; it does not claim node-content edits, media generation, visual-quality review, or publishing.
- When a request mixes arrangement with content, generation, research, or publishing, layout success does not satisfy those other goals; inspect their separate results and approval steps.
Session controls
- Start a new conversation without erasing older history.
- Choose the orchestrator model and available thinking level.
- Keep a turn running while the bottom bar is collapsed or switched back to Make.
- Use the activity label to see whether the agent is planning, building, or waiting.
Good plan boundaries
A useful plan names the graph stages, inputs, expected outputs, approval points, and which existing nodes/Brain records to reuse. Editing the plan is cheaper than undoing a large generated graph.
Layout-only completion
- For a request whose entire goal is arranging existing Flow cards, successful arrangement plus a current graph read-back can be enough for the Agent to acknowledge the layout change.
- A layout-only acknowledgement does not mean document content was changed, a design was visually reviewed, media was generated, or an artifact was published.
- If the request includes any other work, the Agent keeps those obligations and must not treat card arrangement as completion of the full request.
Tips
- Select or name existing nodes when the agent should modify rather than duplicate them.
- Ask the agent to stop before expensive runs when you want to inspect the graph first.
Limitations and important notes
- There is no MCP tool to open the in-product Flowgen Agent panel.
- Tool availability depends on the active Flowgen surface and permissions.
- An external MCP host uses its own model and Flowgen tools; its chat history is not automatically the same as the in-app Agent conversation.
- A layout-only completion confirms canvas arrangement under the current graph scope; it is not evidence of content editing, generation, visual QA, or delivery.