AI agents
Configure an AI agent node: its model, its tools (HTTP, MCP, and code), and the reasoning loop.
The AI agent node is an LLM that reasons over its input and calls tools in a loop until it has an answer. You give it a model and a set of tools; it decides which tools to call and when.
Two simpler AI nodes exist alongside it: LLM call (a single prompt to a model, no tools) and MCP tool (call one named tool on a remote MCP server). This page covers the full agent.
How it's wired
An AI agent is configured with sub-nodes attached to it, not as separate steps in the main flow. An agent needs exactly one model, and any number of tools.
Model
The provider, model, and sampling settings the agent reasons with.
Tools
HTTP, MCP, and code tools the agent can call.
Loop
The agent reasons, calls tools, and repeats up to a set number of iterations.
Model
Attach one model sub-node. It sets:
- Provider and model: any model you've added on the Models page (OpenAI, Anthropic, Gemini, and the rest).
- Sampling: temperature, max tokens, top-p, and top-k.
You add and manage the underlying credentials once under Models; here you just pick which one this agent uses.
Tools
Attach any number of tool sub-nodes. The agent is told what each tool does and calls them as needed:
| Tool | What the agent can do |
|---|---|
| HTTP request tool | Call an API. Supports auth (none, basic, bearer, header, query, or OAuth2 client-credentials). |
| MCP client tool | Use tools from a remote MCP server. Pick the subset of that server's tools to expose. |
| Code tool | Run a Python function the agent can call with arguments. |
Group related tools together so the agent has a clear toolset.
The loop
When the agent runs, it reasons with the model, optionally calls one or more tools, reads the results, and continues until it produces a final answer or hits its max iterations (8 by default). The whole thing is one node in your workflow; its output flows to the next node like any other.
When to use which AI node
- AI agent: the task needs the model to decide and act across several steps or tools.
- LLM call: you just need one prompt-to-text transform (summarize, classify, rewrite).
- MCP tool: you want to call one specific tool, no reasoning loop.
Next
- Models to add a provider
- Nodes for the rest of the node types
- Integrations for app actions