Chat & AI Agents

Execute Large Language Model (LLM) requests and build AI agents that can use tools to perform actions.

Capabilities

Feature	Description
Basic Chat	Simple LLM conversations with OpenAI, Anthropic, Gemini, and more
Tool Calling	AI agents that execute DAG workflows as tools during conversations

Quick Start

Simple Chat

steps:
  - type: chat
    llm:
      provider: openai
      model: gpt-4o
    messages:
      - role: user
        content: "What is 2+2?"
    output: ANSWER

AI Agent with Tools

# Main DAG that uses tools
steps:
  - type: chat
    llm:
      provider: anthropic
      model: claude-sonnet-4-20250514
      tools:
        - calculator
    messages:
      - role: user
        content: "What is 15 times 23?"

---
# Tool DAG definition
name: calculator
description: "Perform basic arithmetic operations"
defaultParams: "operation a b"

steps:
  - name: calculate
    script: |
      case "$1" in
        multiply) echo $(($2 * $3)) ;;
        *) echo "Unknown operation" ;;
      esac
    output: RESULT

Supported Providers

Provider	Environment Variable	Models
openai	OPENAI_API_KEY	GPT-3.5, GPT-4, GPT-4o
anthropic	ANTHROPIC_API_KEY	Claude 3, Claude 4, Claude Sonnet
gemini	GOOGLE_API_KEY	Gemini 1.5, Gemini 2
openrouter	OPENROUTER_API_KEY	100+ models via OpenRouter
local	(none)	Ollama, vLLM, LM Studio (OpenAI-compatible)

Key Features

• Multi-Provider Support - Switch between OpenAI, Anthropic, Gemini, and local models
• Conversation History - Automatic message inheritance between dependent steps
• Extended Thinking - Enable reasoning mode for complex tasks (Anthropic, OpenAI, Gemini)
• Secret Masking - Automatic masking of sensitive values before sending to LLM
• Tool Calling - Build AI agents that execute workflows as function calls
• Variable Substitution - Use ${VAR} in messages for dynamic content
• DAG-Level Defaults - Share LLM configuration across multiple steps

Configuration

Basic LLM Config

Field	Type	Default	Description
provider	string	required	LLM provider (openai, anthropic, gemini, etc.)
model	string	required	Model identifier
temperature	float	provider default	Response randomness (0.0-2.0)
maxTokens	int	provider default	Maximum tokens to generate
stream	bool	true	Stream response tokens

Tool Calling Config

Field	Type	Default	Description
tools	string[]	[]	DAG names to expose as callable tools
maxToolIterations	int	10	Maximum tool calling loops

Examples

Multi-turn Conversation

type: graph

steps:
  - name: setup
    type: chat
    llm:
      provider: openai
      model: gpt-4o
      system: "You are a math tutor."
    messages:
      - role: user
        content: "What is 2+2?"

  - name: followup
    depends: [setup]
    type: chat
    llm:
      provider: openai
      model: gpt-4o
    messages:
      - role: user
        content: "Now multiply that by 3."

The followup step automatically inherits conversation history from setup.

DAG-Level Configuration

# Share LLM config across all chat steps
llm:
  provider: anthropic
  model: claude-sonnet-4-20250514
  temperature: 0.7

steps:
  - type: chat
    messages:
      - role: user
        content: "Explain quantum computing"

  - type: chat
    messages:
      - role: user
        content: "Now explain it to a 5-year-old"

Both steps inherit the DAG-level LLM configuration.

See Also

• Basic Chat - Complete reference for simple LLM conversations
• Tool Calling - Build AI agents with DAG-based tools
• Secrets - Secure handling of API keys
• Variables - Dynamic content in messages