Automations

AI LLM Node (Claude / GPT / Gemini)

8 min readnode-ai-llm

Call any major AI model — Anthropic Claude, OpenAI GPT-4o, or Google Gemini — from an automation step. Supports system prompts, temperature, and variable output.

AI LLM Node

The AI LLM node sends a prompt to a large language model and stores the response as a variable. It supports three providers — Anthropic Claude, OpenAI GPT, and Google Gemini — configurable per node so you can mix providers within the same automation.

---

Fields
FieldTypeRequiredDescription
providerenumYesclaudeopenaigemini
modelstringNoSpecific model ID. Defaults: Claude → claude-sonnet-4-6, GPT → gpt-4o, Gemini → gemini-1.5-pro
systemstringNoSystem / persona prompt. Supports {{variables}}
promptstringYesThe user-turn prompt. Supports {{variables}} and {{lead.*}}
maxTokensintegerNoMax output tokens (default: 1024)
temperaturefloatNo0.0–1.0 (default: 0.7). Lower = more deterministic
outputVarstringNoVariable name to store the response text (default: llmResponse)

---

Output variables
VariableDescription
{{variables.llmResponse}}The model's text response (or your custom outputVar)

---

Example: classify inbound email

json { "provider": "claude", "system": "You are a lead qualification assistant. Respond with a single word only.", "prompt": "Classify this email as one of: hot / warm / cold / spam.\n\nEmail: {{email.bodyText}}", "outputVar": "leadTemperature", "maxTokens": 10, "temperature": 0.1 }

Then use a logic.if node downstream: {{variables.leadTemperature}} equals hot.

---

Example: draft a personalised reply

json { "provider": "openai", "model": "gpt-4o", "system": "You are {{lead.businessName}}'s account manager. Write in a friendly, concise tone.", "prompt": "Draft a 3-sentence reply to this inbound message. Propose a 15-minute call.\n\nMessage: {{email.bodyText}}", "outputVar": "draftReply", "maxTokens": 300 }

Pipe {{variables.draftReply}} into a Send Email node's body field.

---

Provider comparison
ProviderBest forSpeedCost
ClaudeNuanced reasoning, long documents, codeFastMedium
GPT-4oBroad capability, multimodal, function callingFastMedium
Gemini 1.5 ProVery long contexts (1M tokens), multilingualMediumLow

---

Tips
  • Set temperature: 0.1 for classification tasks — you want consistent labels, not creative variation
  • Use the system field to fix the persona and response format; it saves tokens on the prompt itself
  • Chain two LLM nodes: first extract structured data (ai.extract), then generate copy using the extracted values
  • The node stores raw text. If you need JSON, add "Respond with valid JSON only". to your system prompt, then parse with a data.code node