How to Build an AI Workflow with n8n

By the end of this guide you’ll have a working AI workflow in n8n: a chat window connected to an AI Agent that answers questions - and reaches live Wikipedia when it needs facts it doesn’t know. You’ll wire it with the current official node names, verify each step in the execution log, dodge the real error messages, and learn the export/import habits that make these workflows shareable.

If you already know what AI agents are or how MCP connects AI tools, you’ll see n8n as the orchestration layer that binds them - its AI Agent node is the same pattern, pointed at real systems.

What an AI workflow is in n8n

In the builder you’ll see three kinds of nodes:

• Root nodes - where execution starts (chat trigger, webhook, schedule).
• Cluster root nodes - the AI Agent: it makes decisions and orchestrates the rest.
• Sub-nodes - the model, the tools, the memory, the output parser. They attach to the agent as connected “clusters”, not as separate pipeline steps.

The workflow this guide builds: Chat Trigger → AI Agent (OpenAI Chat Model + Wikipedia tool). The agent, at runtime, decides when to search - exactly the pattern your own agent follows, but without writing code.

Step 1: Pick your instance: n8n Cloud or self-hosted

The official docs line on self-hosting is worth quoting: “Self-hosting n8n requires technical expertise” - the built-in Docker install docs are even tagged as outdated in favor of a Docker Compose setup. For this guide, n8n Cloud’s free trial is the recommended path (sign up at app.n8n.cloud/register).

Self-hosting on the official image (uses the container’s demos/hello-world page at http://localhost:5678):

docker run -it --rm \
  -p 5678:5678 \
  -e N8N_SECURE_COOKIE=false \
  -v n8n_data:/home/node/.n8n \
  docker.n8n.io/n8nio/n8n

Step 2: Create the workflow (import it, don’t rebuild it)

Start from Workflows → Add workflow (blank), or skip the dragging entirely - import the copy-ready JSON below via the workflow’s … menu → Import from file (the official export/import path). It contains all four nodes with the exact current type identifiers:

{
  "name": "AI Workflow - Wikipedia Research Assistant",
  "nodes": [
    {
      "parameters": {
        "public": true,
        "options": {}
      },
      "id": "chat-trigger",
      "name": "Chat Trigger",
      "type": "n8n-nodes-langchain.chatTrigger",
      "typeVersion": 1.1,
      "position": [0, 0],
      "webhookId": "CHAT_WEBHOOK_ID"
    },
    {
      "parameters": {
        "promptType": "define",
        "text": "You are a research assistant. Answer questions from the user. If you need current or factual details, search Wikipedia with the tool before answering. Keep answers concise (max 3 paragraphs) and always cite the entry you used.",
        "options": {}
      },
      "id": "ai-agent",
      "name": "AI Agent",
      "type": "n8n-nodes-langchain.agent",
      "typeVersion": 2,
      "position": [220, 0]
    },
    {
      "parameters": {
        "options": {},
        "model": "gpt-4o-mini"
      },
      "id": "openai-model",
      "name": "OpenAI Chat Model",
      "type": "n8n-nodes-langchain.lmChatOpenAI",
      "typeVersion": 1,
      "position": [220, 220],
      "credentials": {
        "openAiApi": {
          "id": "YOUR_OPENAI_CREDENTIAL_ID",
          "name": "OpenAI account"
        }
      }
    },
    {
      "parameters": {
        "options": {}
      },
      "id": "wikipedia-tool",
      "name": "Wikipedia",
      "type": "n8n-nodes-langchain.toolWikipedia",
      "typeVersion": 1,
      "position": [440, 220]
    }
  ],
  "connections": {
    "Chat Trigger": {
      "main": [[{ "node": "AI Agent", "type": "main", "index": 0 }]]
    },
    "AI Agent": {
      "ai_languageModel": [[{ "node": "OpenAI Chat Model", "type": "ai_languageModel", "index": 0 }]],
      "ai_tool": [[{ "node": "Wikipedia", "type": "ai_tool", "index": 0 }]]
    }
  },
  "settings": { "executionOrder": "v1" }
}

If you prefer to build by drag-and-drop instead of importing: add Chat Trigger (root, core-nodes), add AI Agent (cluster root), connect trigger → agent, then attach sub-nodes: OpenAI Chat Model (models → chat model) and Wikipedia (tools → Wikipedia). Wire the connectors and you’re done - each sub-node needs only its connector line, not a value.

Step 3: Set up the OpenAI credential

1. In the OpenAI Chat Model node, open Credential and choose Create new credential (type OpenAI API).
2. Paste your OpenAI API key - the only required field. Organization ID is optional per the official credentials docs, only needed if your account belongs to multiple organizations.
3. Save, then check the Model dropdown - models load dynamically from the API once a valid credential exists; pick gpt-4o-mini or whatever your account has access to.
4. Leave the default model options alone for now. There is no required parameter beyond the credential.

The system prompt in the agent node does the steering: it tells the model to search Wikipedia before answering. Keep it short - the tools you connect do the heavy lifting.

Step 4: Run it - and watch it think

1. Execute the workflow (the Test workflow dialog opens the chat window embedded in the current editor).
2. Send a question that demands live knowledge: “When was Portugal’s 25 April Bridge built - and how long is it?”
3. Answer will be composed by the model - and the executions log (bottom panel) shows every step: the trigger, the agent, two OpenAI calls (tool round-trip), and the Wikipedia node.

Two behaviors to expect, both official:

• Every chat message = one workflow execution. n8n bills per execution on Cloud plans, and large prompts add model API cost. This is the biggest difference between a chat workflow and a plain webhook.
• Chat sessions persist via memory. By default the agent keeps per-session context; adding a Simple Memory sub-node keeps history across messages (it’s what the “Load Previous Session” option in the trigger’s session settings uses).

Step 5: Verify it worked, then make it yours

Verification checklist (official, in order):

1. Execution was successful - no red nodes in the executions log; the Chat Trigger shows 200 status.
2. The tool was used - the Wikipedia node has an execution entry with query/output (subset of the answer’s facts).
3. The answer cites the entry - the system prompt’s instruction surfaced in the reply.
4. Duplicate check - “same question twice”: memory sub-node not attached → identical answers; attached → context-aware follow-ups.

Then remix it:

• Add tools from the official tool sub-node list - calculator, code, SEARXNG web search, WolframAlpha, HTTP Request over toolMCP (any MCP server you set up with our MCP guide), plus memory and output parsers.
• Most tools need credentials (Wikipedia doesn’t).
• Export your work: … menu → Download file writes JSON; Import from URL/File restores it. Copy-paste nodes between workflows (Ctrl/Cmd+C → V) also works on the canvas.

Troubleshooting

You did it

• You chose Cloud vs self-hosted with the cost table in mind
• The workflow runs end-to-end from the Chat Trigger’s test chat
• OpenAI credential attached - no auth error on execute
• One answer in the execution log proves the Wikipedia tool fired
• You can export the workflow to JSON and re-import it cleanly

Next: connect your workflow to the outside AI universe - point an MCP client at n8n’s built-in MCP server so Claude can drive your workflows, or deep-dive into what makes agents tick under the hood.