Agent
For a ready-made starting workflow, see Templates.
Start with Agent
The Agent can build and edit a workflow from your description. It uses Claude Code or Codex on your computer. These tools use their own account sign-in, separate from generation-provider keys.
- Select Start with Agent on the Welcome screen, or open Agent on the canvas.
- Choose Claude Code or Codex.
- If the selected tool is missing, use the installation command shown in the panel.
- Select Open Claude Code sign-in or Sign in with ChatGPT.
- Complete the account sign-in in your browser.
- Describe the inputs, steps, and outputs you want.
- Check the proposed nodes, models, prompts, and noodles before generation.
Claude Code needs a paid Claude subscription. Codex needs a ChatGPT plan that provides Codex access. The Agent requires confirmed subscription billing; provider API keys do not enable it. Generation nodes still need their selected providers' keys. The app stops Agent work at the detected plan limit instead of falling back to extra usage or credits.
For example:
Use a portrait image as a reference.
Create three background descriptions with an LLM.
Generate one portrait variation for each description.
Collect the results in an Output Gallery.An LLM is a large language model that reads and generates text. AI-generated workflows can need changes to model choices, prompts, or sockets. You can edit those parts with the same controls as a manual workflow.
Choose a model and manage a conversation
| Control | What it does |
|---|---|
| Agent and model menu | Selects Claude Code or Codex and a model available through that tool |
| Thinking effort | Sets the effort level for a model that supports it |
| Enter | Sends a message, or queues it while a reply runs |
| Shift + Enter | Adds a new line |
| Stop | Stops the current Agent turn |
| New chat | Starts a conversation and clears queued messages |
The model menu can show a built-in list if the tool's model list is unavailable. You cannot switch tools while a reply runs. Closing the Agent window hides it; the current turn continues.
Messages sent during a turn enter a queue above the composer. You can edit or remove queued messages. After a successful turn, the Agent processes them in order. After a stopped or failed turn, the queue waits for you to choose Send now. Switching tools clears the queue.
Completed conversations appear in local history. History stays in that browser profile and holds up to 50 conversations within a storage limit. It is separate from the saved workflow file.
Undo and refine changes
Use Cmd/Ctrl + Z to undo Agent changes.
Each applied batch creates an undo step, so a turn with several batches can need several undo steps.
If a workflow differs from your intent, describe the missing step in a follow-up message. The Agent carries earlier preferences through the current conversation. A new chat does not provide persistent project instructions.
If a model needs a key, add that provider in Settings or choose a configured model.
If sockets do not match, select a compatible model and reconnect the affected noodles.
Save a useful result with Cmd/Ctrl + S or Nb → Save project.
Look up prompting tips
- Select one generation node with a model.
- Select Look up prompting tips in the Agent.
- Let the research turn finish and save its notes.
- Use Refresh for new research, or Forget to remove the saved tips.
These notes describe how to prompt that model.
They persist across chats and projects on this computer.
The default notes folder is ~/.node-banana/prompt-notes.
They are separate from conversation history and your workflow prompts.
Next steps
- For sign-in errors, see Troubleshooting.
- See Best practices for batch and reuse patterns.