Core Nodes

start

Every flow must have exactly one start node. It defines the entry point and trigger configuration.

{
  "id": "start-1",
  "type": "start",
  "position": { "x": 0, "y": 0 },
  "data": {
    "type": "start",
    "inputVariable": "input",
    "welcomeMessage": "Hello! How can I help you?",
    "trigger": {
      "type": "manual"
    }
  }
}

Field	Type	Default	Description
inputVariable	string	"input"	Variable name receiving user input
welcomeMessage	string	—	Greeting shown before first input
trigger	TriggerConfig	—	How the flow is activated

agent

The core workhorse — sends messages to an LLM with a system prompt, tools, and context.

{
  "id": "agent-1",
  "type": "agent",
  "position": { "x": 0, "y": 200 },
  "data": {
    "type": "agent",
    "modelId": "openai/gpt-4o-mini",
    "systemPrompt": "You are a helpful assistant.",
    "temperature": 0.7,
    "maxTokens": 4096,
    "responseTone": "professional",
    "mcpTools": [],
    "brandMemoryCollectionIds": [],
    "outputVariable": "response",
    "streaming": true,
    "chatHistoryLimit": 20,
    "maxToolIterations": 10
  }
}

Field	Type	Default	Description
modelId	string	—	Model ID from OpenRouter (e.g., openai/gpt-4o)
systemPrompt	string	—	Instructions for the agent
temperature	number	0.7	Response randomness (0–2)
maxTokens	number	4096	Maximum response length
responseTone	string	—	Tone: professional, casual, friendly, formal, technical, creative
mcpTools	AgentMCPServerConfig[]	[]	MCP tool server configurations
maxToolIterations	number	10	Max tool-call loop iterations (max 50)
brandMemoryCollectionIds	string[]	[]	RAG collection IDs
outputVariable	string	—	Variable to store the response
streaming	boolean	true	Stream the response
chatHistoryLimit	number	—	Limit chat history messages (0 = none)
routingStrategy	RoutingStrategy	—	Dynamic model selection
fileAccess	FileAccessConfig	—	File access permissions

MCP Tool Configuration

{
  "mcpTools": [
    {
      "catalogId": "memory",
      "name": "Memory",
      "transport": "streamable-http",
      "url": "http://mcp-memory:8001/mcp",
      "authType": "none",
      "selectedTools": ["create_memory", "search_memory"],
      "timeout": 30000
    }
  ]
}

classify

Routes input to different branches based on LLM classification.

{
  "id": "classify-1",
  "type": "classify",
  "data": {
    "type": "classify",
    "modelId": "openai/gpt-4o-mini",
    "categories": [
      { "id": "question", "name": "Question", "description": "User is asking a question" },
      { "id": "complaint", "name": "Complaint", "description": "User has a complaint" },
      { "id": "praise", "name": "Praise", "description": "User is giving positive feedback" }
    ],
    "instructions": "Classify the user's intent",
    "confidenceThreshold": 0.7
  }
}

Each category ID becomes a source handle for outgoing edges.

end

Terminates the flow and returns output.

{
  "id": "end-1",
  "type": "end",
  "data": {
    "type": "end",
    "status": "success",
    "outputVariable": "response",
    "targetPlatform": "twitter",
    "platformConnectionId": "conn-123"
  }
}

Field	Type	Description
status	success | error | cancelled	Completion status
outputVariable	string	Variable with the final output
message	string	Static message (if no variable)
targetPlatform	string	Platform to post response to

note

Non-executable annotation node for documentation on the canvas.

{
  "id": "note-1",
  "type": "note",
  "data": {
    "type": "note",
    "content": "This section handles user authentication",
    "backgroundColor": "#1a1a2e",
    "width": 300,
    "height": 150
  }
}

agent_ref

Delegates to a sub-agent by slug or ID.

{
  "id": "ref-1",
  "type": "agent_ref",
  "data": {
    "type": "agent_ref",
    "agentRef": "research-assistant",
    "refType": "slug",
    "inputMapping": { "message": "{{input.query}}" },
    "outputVariable": "research_result",
    "inheritContext": true,
    "timeout": 60000
  }
}

agent_pool

Runs multiple agents and combines results.

{
  "id": "pool-1",
  "type": "agent_pool",
  "data": {
    "type": "agent_pool",
    "agents": [
      { "ref": "writer", "refType": "slug", "weight": 1 },
      { "ref": "editor", "refType": "slug", "weight": 1 }
    ],
    "strategy": "best_of",
    "evaluationPrompt": "Pick the best written response",
    "evaluatorModelId": "openai/gpt-4o",
    "maxConcurrency": 3,
    "agentTimeout": 30000
  }
}

Strategies: best_of, aggregate, vote, fastest

map_reduce

Processes an array of items in parallel with a map agent, then reduces results.

{
  "id": "mr-1",
  "type": "map_reduce",
  "data": {
    "type": "map_reduce",
    "inputVariable": "articles",
    "mapAgentRef": "summarizer",
    "mapAgentRefType": "slug",
    "itemTemplate": "Summarize: {{item.title}}\n{{item.content}}",
    "reduceStrategy": "concatenate",
    "maxConcurrency": 5
  }
}

Reduce strategies: concatenate, merge, custom