AI Agents — Node reference

The modular AI agent family is six nodes you’ll find in the palette under the AI / AGENTS category. This page describes them one by one with their fields.

Note

Every sub-node (brain, tool, RAG, skill) carries two common fields: role (its function inside the agent) and attach_to (alias of the AI Agent it belongs to). In the editor they are filled in automatically when you connect to a port; you can edit them by hand if you prefer.

AI Agent (orchestrator)

SmartToySharp · The central node. Runs in the flow, discovers the connected pieces and executes the reasoning loop with function calling.

Field	Type	Required	Description
system_prompt	textarea	No	Base instructions for the agent (personality, rules, goal).
input	text	No	User message/input. If left empty, it uses the data arriving from the previous node (data.input, data.message or the whole data).
temperature	number	No	Answer creativity. 0 = deterministic, 1 = creative. Default 0.7.
max_iterations	number	No	Max rounds of the tool loop (request tool → run → repeat). Default 5.

Input ports (bottom edge): brain (1), tools (N), rag (N), skills (N).

Output: data.response (final text) and data.iterations (how many tool rounds it ran).

Brain (LLM)

PsychologySharp · The model that reasons. Connects to the brain port. One brain per agent.

Field	Type	Required	Description
provider	select	Yes	LLM provider. See the providers table below.
model	aiModelSelector	Yes	Model, depending on the chosen provider (e.g. gpt-4o, claude-sonnet-5, mistral-large-latest).
credentials_id	credentials	Yes	Provider credential.
attach_to	text	Yes*	Agent alias. *Filled automatically when connecting to the brain port.

Supported providers

provider	Service	API
openai	OpenAI (GPT)	native
anthropic	Anthropic (Claude)	native
google	Google (Gemini)	native
deepseek	DeepSeek	OpenAI-compatible
qwen	Alibaba Qwen	OpenAI-compatible
groq / llama	Groq (Llama)	OpenAI-compatible
mistral	Mistral AI	OpenAI-compatible
kimi	Moonshot (Kimi)	OpenAI-compatible
openrouter	OpenRouter	OpenAI-compatible
together	Together AI	OpenAI-compatible
local	Ollama / vLLM	OpenAI-compatible (base URL in the credential)

Tip

Most providers expose an OpenAI-compatible API, so they are handled with a single client by changing the base URL. For local LLM (Ollama/vLLM), the base URL comes from the local credential (e.g. http://localhost:11434/v1).

Tool

BuildSharp · Declares a capability the agent can invoke. Connects to the tools port. You can connect several.

Field	Type	Required	Description
tool_name	text	Yes	The function name the LLM will see (e.g. get_weather). No spaces.
tool_description	textarea	Yes	What it does. The LLM decides when to use it by reading this — be clear and specific.
tool_schema	json	No	JSON schema of the parameters (JSON Schema format). If omitted, the tool receives no arguments.
tool_kind	select	Yes	What the tool runs: node, workflow or registered.
node_type	text	Conditional	If tool_kind = node: the node type to run (e.g. openaiImages, http, finnegansGet).
node_config	json	No	Base config of the node (merged with the arguments the LLM decides).
workflow_id	text	Conditional	If tool_kind = workflow: the id of the workflow the tool wraps.
tool_id	text	Conditional	If tool_kind = registered: the id in the agent_tools table.
attach_to	text	Yes*	Agent alias.

The three tool kinds

• node — any of the 600+ catalog nodes becomes a capability of the agent (generate images, call an API, query an ERP, read an Excel file…).
• workflow — a whole workflow is exposed as a single tool (useful for multi-step processes).
• registered — a tool already defined in agent_tools (compatibility with the previous tools system).

RAG (retrieve)

TravelExploreSharp · Retrieves the relevant chunks from the vector store for a query. Dual use: as a standalone flow node, or connected to the agent’s rag port (injects context before answering).

Field	Type	Required	Description
namespace	text	Yes	Namespace to search in (must match the one used when indexing).
top_k	number	No	How many chunks to retrieve. Default 4.
provider	select	Yes	Embeddings provider (openai, local, together, mistral).
model	text	Yes	Embeddings model (e.g. text-embedding-3-small).
credentials_id	credentials	Yes	Embeddings provider credential.
attach_to	text	No*	Agent alias (only if used as a RAG sub-node).

Caution

The embeddings model when retrieving must be the same one used when indexing, and so must the namespace. If they don’t match, the search returns poor or empty results.

RAG (ingest)

StorageSharp · Loads text/documents into the vector store (splits into chunks, computes embeddings and stores them). It’s a flow node (it does not connect to the agent); you use it in an ingestion workflow.

Field	Type	Required	Description
namespace	text	Yes	Namespace to store into.
provider	select	Yes	Embeddings provider.
model	text	Yes	Embeddings model (same one you’ll use when retrieving).
credentials_id	credentials	Yes	Provider credential.
chunk_size	number	No	Size of each chunk in characters. Default 1000.
overlap	number	No	Overlap between chunks. Default 150.

Input: data.text (a single text) or data.documents (an array [{ text, id, metadata }]).

Skill

AutoAwesomeSharp · A package of instructions + tools the agent activates only when the request is relevant. Connects to the skills port. You can connect several.

Field	Type	Required	Description
skill_name	text	Yes	Skill name (e.g. billing).
description	textarea	Yes	When to use it. Compared against the user request to decide whether to activate.
keywords	text	No	Keywords (CSV) that trigger the skill (e.g. invoice, receipt, VAT).
prompt_fragment	textarea	No	Instructions injected into the system prompt when the skill activates.
enables_tools	text	No	Tool names (CSV) this skill enables. See the gating box.
attach_to	text	Yes*	Agent alias.

Tool gating by skill

If a tool appears in some skill’s enables_tools, that tool stays hidden until the corresponding skill activates. Tools not referenced by any skill are always available. This way you avoid exposing the LLM to tools that aren’t relevant.

Next step

With the nodes clear, continue with How to combine the nodes.