Skip to main content
sndbox exposes AI in three deliberately separate places. All three use an AI connection, but they affect a workflow differently.

​
Draft a graph

  1. Open a workflow and choose the sparkle button in the top toolbar to open AI builder.
  2. Select a connected model.
  3. Describe the trigger, steps, data flow, and desired result. Include constraints such as “draft the email but do not send it.”
  4. Review the proposed node count, connections, validation issues, and summary.
  5. Choose Apply workflow only when the proposal matches your intent.
  6. Configure any deliberately blank credentials, paths, URLs, recipients, or selectors, then validate and run the workflow yourself.
The builder returns a replacement graph, not an execution. It can create built-in nodes, control edges, typed input bindings, and complete Code node source blocks. It understands the capabilities and outputs of the built-in node catalogue, so a request such as an uptime monitor uses HTTP, state, condition, and notification nodes for the monitoring work while Code and Web Builder nodes create the interface. Web Builder’s three typed inputs are repaired from the matching Code-node languages when a model returns a generic connection. While a draft is in progress, the chat shows muted activity stages such as Planning Draft, Investigating current workflow, and Writing Code. These are concise operation indicators rather than hidden model reasoning.
Applying a proposal changes the editor draft. AI does not grant permissions, approve risky actions, invent secret values, or start the workflow. Review side effects and permission changes exactly as you would for a manually built graph.

​
Use AI during a run

Add an AI node when the next workflow step depends on a live answer. Select a connection, enter or map the instruction, optionally refine the system instruction, and choose the maximum output tokens, creativity, and timeout. The workflow waits until one of these events occurs:
  • The provider returns a response.
  • The configured timeout is reached.
  • The execution is cancelled.
  • The provider rejects the request or the connection is unavailable.
Map response into a later text-compatible field. usage contains provider-reported token information when available, and model records the model ID used for the response. The default timeout is 90 seconds; the runtime accepts 1–300 seconds.

​
Write code with AI

Open a Code node, choose Edit code, then choose Code with AI. A coding chat opens to the right of the editor. Select a connection and describe the complete file or requested change. The current language and unsaved editor draft are included as context on every turn. Each completed response replaces only the unsaved draft and remains in the chat so you can refine it. The Problems panel immediately checks the returned source. Choose Save code or press Ctrl+S / Cmd+S to commit it to the node; cancel the dialog to keep the previously saved source. AI-generated code is untrusted code. Inspect it before saving, and review command-execution permission before using Python or JavaScript in run mode. See Code and Web Builder for execution and localhost details.

​
Prompting tips

  • Name the trigger and final outcome.
  • State which external actions must remain drafts or disabled.
  • Describe the expected shape of data between nodes.
  • For code, request a complete file and state the language, inputs, output format, and error behavior.
  • For a site, describe the HTML structure, interactions, styles, and the data each code block should contain.