LangChain Code node examples: a chain and a Wikipedia tool

LangChain Code node examples: a chain and a Wikipedia tool

An n8n workflow with two small demos of n8n's LangChain Code node. In the first, hand-written code builds a prompt for
"Tell me a joke", pipes it into an OpenAI chat model and returns the answer. In the second, an AI Agent answers "What
year was Einstein born?" with a Wikipedia tool that a second Code node supplies.

What is inside

• workflow.json: the workflow as n8n exports it, ten nodes. Eight do the work and two are sticky notes that name the
  demos ("Self-coded LLM Chain Node" and "Self-coded Tool Node"). It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The first demo's model was n8n's OpenAI Model node, a completions model. With no model set it used
  gpt-3.5-turbo-instruct, which OpenAI shut down on 28 September 2026. The replacement OpenAI names,
  gpt-5.6-terra, does not take the legacy completions endpoint. The node is now an OpenAI Chat Model, version 1.2,
  named OpenAI Chat Model, with gpt-5.6-terra in its ID field.
• A chat model returns a message, not plain text, so the chain's code now ends with (await chain.invoke()).content.
  The output is the reply's text, as before.
• The chain imported langchain/prompts and the tool imported langchain/tools. n8n's own starter code for the node,
  read on 12 November 2026, imports @langchain/core/prompts and @langchain/community/tools/wikipedia_query_run, so
  the copy uses those paths. Its loader rewrites the old langchain/tools path but not langchain/prompts.
• Both model nodes were linked to a credential in their author's n8n. The links are gone, so select your own.
• The workflow's tag and its ids are gone.
• Nothing else. The questions, the agent and the layout are the file's own.

Before you start

• An n8n instance with the LangChain Code node, on n8n Cloud or your own server.
• An OpenAI API key, for an OpenAI credential. Wikipedia needs no key.
• Some JavaScript, to read the two scripts.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add an OpenAI credential and select it in OpenAI Chat Model and in Chat OpenAI.
3. Open Set and Set1 to read the two questions. Change them if you like.
4. Press Execute workflow. Open Custom - LLM Chain Node for the joke and Agent for the Einstein answer.

How a run goes

1. The manual trigger starts both demos. It feeds Set and Set1, which write one input field each.
2. Custom - LLM Chain Node (LangChain Code, execute mode) has one main input and one required language model input.
   Its code builds a prompt with PromptTemplate.fromTemplate from input, reads the connected model with
   getInputConnectionData('ai_languageModel', 0), pipes the prompt into it and returns an item with output.
3. Agent (AI Agent 1, a conversational agent by default) reads its question from input. It uses Chat OpenAI, whose
   model is left at the node's default, and Custom - Wikipedia.
4. Custom - Wikipedia (LangChain Code, supply data mode, an AI tool output) logs a line and returns a
   WikipediaQueryRun object. The agent calls it when it wants to look something up.

Change the look

• Questions: the input value in Set and Set1.
• Chain: the code in Custom - LLM Chain Node. Chain more steps after the prompt, or parse the answer.
• Tool: the code in Custom - Wikipedia. Return any LangChain tool object, such as one that calls your own search.
• Models: OpenAI Chat Model for the chain and Chat OpenAI for the agent. The chain reads whichever model is
  connected, so swapping the node needs no code change.
• Debugging: console.log inside a Code node writes to the browser's console.

Good to know

• The chain treats the question as a prompt template, so a question with { and } in it is read as a variable.
• Chat OpenAI leaves its model at the node's default, gpt-5-mini. OpenAI's deprecations page, read on 12 November
  2026, shuts that model down on 11 December 2026, with gpt-5.6-terra as the replacement.
• A model can answer the Einstein question from memory and skip the tool. Watch the console for the line the tool
  logs to see whether it ran.
• OpenAI bills one call for the chain and at least one for the agent.
• Both questions go to OpenAI, and the agent's search terms go to Wikipedia.
• The file holds no credential and no key.