Node reference

Agent node: Claude with tools and structured output

agent · Updated September 17, 2026

The Agent node is an LLM call with tool calling and structured output. It uses your Anthropic credential, runs a tool loop when tools are attached, and streams tokens to the canvas while it works.

When to use it

Classifying, extracting or summarising text.
Deciding what to do next, with structured output that an If or Switch node can branch on.
Research or lookups, with HTTP Request nodes attached as tools.

Ports

Inputs: intools
Outputs: outerror

Configuration

Credentialcredential_idrequired: credential
Modelmodelrequired: stringPick a model, or choose Other to enter any model ID.
System promptsystem_prompt: string
User promptuser_promptrequired: template
Efforteffort: none | low | medium | high | xhigh | max · default "none"None sends no effort, which Haiku and older models need.
Temperaturetemperature: number · default 1Only with Effort set to None. Claude 5 models reject temperature.
max_tokensmax_tokens: integer · default 2048
Max tool roundsmax_iterations: integer · default 6
Structured output (JSON Schema)output_schema: object

Output

Fires out with:

text: the final answer.
structured: an object matching the output schema, when one is set.
tool_calls, usage (input and output tokens) and stop_reason.

Example

User prompt

Classify this ticket: {{ trigger.message }}

Structured output (JSON Schema)

{
  "type": "object",
  "properties": { "urgency": { "type": "string", "enum": ["low", "normal", "high"] } },
  "required": ["urgency"]
}

Good to know

Effort defaults to None, which sends no effort setting. Haiku and older models require that, and only then is Temperature available. Claude 5 models reject temperature.
The output schema is made strict automatically (additionalProperties: false on every object), as the API requires.
A failing tool call returns an error result to the model instead of failing the node. Max tool rounds (6 by default) bounds the loop.
Pick a model from the list, or choose Other and enter any model ID.

Build it on the canvas

Create a free account and run your first workflow in the browser, or run Novalink on your own machine.

Get started Self-host it