Use the model selected in Claude, Codex, or another MCP host as the single reasoning agent while it calls flowmo and Flowgen tools directly.
Where: Connected MCP host → f0 and flowgen_* tools
Key ideas
- The host is the agent: The external MCP host owns the conversation, reasoning, model choice, and orchestration. It should call first-class flowgen_* tools rather than forwarding the request to the in-app Flowgen Agent.
- Shared execution layer: External MCP tools and the in-app assistant use the same canonical Flowgen declarations and renderer executors. Direct MCP calls therefore keep node schemas and behavior aligned without creating a second AI loop.
- Separate conversations: An in-app Agent session and an external MCP chat have different histories, attachments, model selections, and billing paths. Canvas state is shared; chat context is not.
Steps
- Keep the user’s request in the MCP host conversation and choose the host model appropriate to the task.
- Call f0_catalog or f0_features for product knowledge and f0_guide when the host has not loaded flowmo’s skills.
- For Flowgen graph work, use first-class flowgen_* tools such as flowgen_get_flow_state, flowgen_create_node, flowgen_connect_nodes, and flowgen_run_flow.
- For editor work, inspect the correct editor/document surface and call the normal f0 tools there.
- Review the canvas result and costs in one agent loop; use the in-app Agent separately only when the user intentionally wants a separate conversation.
A clean division of responsibility
- MCP host: understands the request, plans, selects tools, asks clarifying questions, and narrates changes.
- flowgen_* tools: inspect and mutate the connected graph using the same executor contracts as the app.
- f0 editor tools: inspect and edit the visual document on editor or document surfaces.
- In-app Flowgen Agent: an optional, separate user-facing agent—not a required proxy for MCP.
Why not proxy the prompt
Forwarding an MCP-host prompt into the in-app Agent creates nested reasoning: two model selections, two context windows, two approval paths, and potentially two AI charges. Direct tool calls keep authorship and review legible.
Tips
- Target a specific f0_project_id when several Flowgen tabs are open.
- Read narrowly and inspect before mutating; do not repeatedly send an entire project when one node or element is enough.
Limitations and important notes
- Tool availability depends on the open surface and the user’s permissions.
- The MCP host cannot assume that an in-app chat remembers decisions made externally.