AI Agent

Conversational AI agent with a pluggable model (Ollama local/cloud), pluggable memory (state/file/in-memory), and tools (workflows + MCP). Calls tools in a loop until it can answer.

Finding it in the library

Search the builder's node library for AI Agent (it lives under AI). A single click opens the in-editor docs panel shown here — description, ports, and every property, without leaving the canvas. Double-click (or drag) to add it to the workflow.

Wired up in the builder

AI Agent in a real, runnable flow — captured live from the Studio editor, exactly as it looks on your canvas. The ◈ ports carry the AI Model provider and Agent Memory you wire in. This is the same workflow used for the example input & output below.

How it’s configured

The node’s Configure panel as it opens in the builder when you select the step — every setting laid out with real values. Click any field to edit it.

Ports

Ports are the node’s contract with its neighbours. In the editor a port label renders bold when wired and italic when optional; ◈ ports accept attachment carriers rather than data wires.

How data flows through it

AI Agent consumes the content of the incoming envelope — when it is fed directly by a trigger, the trigger’s wrapper is unwrapped at the node boundary so the node sees the actual data, not the metadata shell. Its output becomes the payload for the next node, while the envelope (trace ids, correlation, binary refs) rides along untouched. In the Runs view you always see the whole envelope for both sides of this node.

Expressions in the config

String-typed properties accept {{ }} expressions evaluated against the incoming item at run time — e.g. {{ $json.customer.email }}. On this node that’s system, memoryEmbedModel, memoryEmbedBaseUrl, memoryEmbedKey, sessionKey, message. JSON- and code-typed fields never interpolate — they are passed through literally.

Build it with AI

Every node in this reference is reachable through Flowdrome’s AI Copilot and the MCP tools — say what you want, and the graph surgery happens server-side. Node types resolve fuzzily, so the catalog label (AI Agent) works as well as the exact type id (ai.agent).

In the Copilot panel (or any connected AI):

add a ai agent node after the trigger

As a step in a create_chain_workflow call:

{"type":"AI Agent","config":{}}

curl -s -X POST http://localhost:48170/mcp -H "content-type: application/json" -d '{ "jsonrpc": "2.0", "id": "1", "method": "tools/call", "params": { "name": "add_node", "arguments": { "workflowId": "<id>", "type": "AI Agent" } } }'

Example input & output

Captured from a real test run of the workflow above — this is what you see in the run data panel after pressing Test workflow.

{
  "kind": "json",
  "contentType": "application/json",
  "payload": {
    "message": "Reply with exactly the word OK."
  },
  "body": {
    "message": "Reply with exactly the word OK."
  }
}

{
  "text": "OK",
  "session": "smoke-session",
  "model": "qwen2.5:0.5b",
  "iterations": 1,
  "toolCalls": [],
  "promptTokens": 29,
  "completionTokens": 2
}

Property reference

Every setting, with its type and default — the same fields shown configured in the panel above.

Using it

The AI Agent is Flowdrome’s flagship AI node: a full agentic loop in one step. Each run it recalls the conversation from memory, sends your message to the model with the system persona and rules, lets the model call tools for up to Max iterations rounds, and emits the final answer (plus the tool-call trace) as the node output.

Model — local or cloud

• Provider = ollama runs against a local Ollama daemon (http://localhost:11434 by default) — no API key, nothing leaves your machine. Any pulled model works (Model = qwen2.5:0.5b is enough to try it).
• Provider = ollama-cloud targets https://ollama.com with bigger hosted models (e.g. gpt-oss:20b). Put the key in a stored credential and reference it from the API key field as ${credential.<name>.apiKey} — never paste a literal key into the workflow.

Memory — three pluggable backends

Memory backend decides where conversation history lives, and Session key decides which conversation a run belongs to (it’s a template — {{ input.sessionId }} gives every customer their own thread; the same key recalls the same memory):

History limit caps how many prior turns are replayed into the prompt.

Tools — functions, workflows, MCP

The Tools list is what makes it an agent. Three kinds, mixable:

• Workflows as tools — { "type": "workflow", "name": "get_order", "description": "Fetch one order by id", "parameters": { … }, "workflowId": "wf_…" } lets the agent invoke your other Flowdrome workflows. The description tells the model when to call it; the parameters schema tells it how. Your existing flows become the agent’s hands.
• Function tools — small declared functions the host resolves directly.
• MCP tools — { "type": "mcp", "command": "…", "args": [ … ] } connects an external Model Context Protocol server and exposes its tools to the agent.

The output’s toolCalls array records every round — which tool, what arguments, what came back — so you can audit exactly how the agent reached its answer.

A real run

The example capture above is a live run against a local Ollama (qwen2.5:0.5b): the output carries text (the reply), session, model, iterations, toolCalls, and token counts. The AI in Flowdrome guide walks the whole setup with screenshots, including the WebSocket order-assistant demo — an agent with three workflow-tools taking orders over a live socket.

Tips

• Start with Max iterations = 1 (no tools) to validate model + prompt, then add tools and raise it.
• Output format = json constrains the final answer to a JSON object — combine with rules that name the fields.
• Keep Temperature at 0 for operational agents; raise it only for creative tasks.
• A blank Base URL is correct for both providers — it defaults per provider.

Related nodes

The rest of the AI group — the same folder you’d scan in the editor’s library.

• AI Classify Classifies JSON content
• AI Prompt Runs one prompt through an AI model and returns the text reply — use it to generate, rewrite, answer, summariz
• AI Embeddings Turns text into embedding vectors using OpenAI-compatible endpoints or a local Ollama daemon
• AI Vector Store Self-contained semantic memory: embeds text internally and inserts, queries (cosine similarity), or clears an
• AI Extract Uses an LLM to pull structured fields (order number, customer, intent, dates, …) out of free text
• AI Eval Score an AI output against assertions: exact, contains, not-contains, regex, JSON-schema, and LLM-judge (a rub
• AI Transcribe Turn speech into text via any OpenAI-compatible /audio/transcriptions endpoint (Whisper, or a local faster-whi
• AI Speak Turn text into spoken audio (TTS) via any OpenAI-compatible /audio/speech endpoint

This page is generated from the node registry by gen-node-docs.mjs on every site build — ports, properties, defaults and visibility rules cannot drift from the code. The screenshots and example data are captured from a live Flowdrome by npm run shots:nodes and npm run gen:examples. The hand-authored notes above are merged in.