n8n - 💡(How to fix) Fix Builder MCP `create_workflow_from_code` returns 500 on any `@n8n/n8n-nodes-langchain.*` root node, despite `validate_workflow` accepting it [1 comments, 2 participants]

Summary

The official n8n Workflow Builder MCP server (public preview, announced 2026-04-29) accepts workflow code containing @n8n/n8n-nodes-langchain.* root nodes via validate_workflow but returns HTTP 500 on create_workflow_from_code and update_workflow for the same code. Reproduced across multiple node types and configurations.

Environment

• MCP server: official n8n Workflow Builder (Cloud)
• SDK: @n8n/workflow-sdk
• Tested via Claude Code MCP integration
• First observed 2026-04-30 with @n8n/n8n-nodes-langchain.openAi v2.1
• Re-tested 2026-05-03 with v2.3 and the AI Agent + Chat Model pattern

Reproduction 1 — minimal v2.3 OpenAI Message a model

import { workflow, trigger, node } from '@n8n/workflow-sdk';

const webhookTrigger = trigger({
  type: 'n8n-nodes-base.webhook',
  version: 2.1,
  config: {
    name: 'Webhook',
    parameters: { httpMethod: 'POST', path: 'mcp-test-min' },
    position: [240, 300],
  },
  output: [{ body: {} }],
});

const messageModel = node({
  type: '@n8n/n8n-nodes-langchain.openAi',
  version: 2.3,
  config: {
    name: 'Message a model',
    parameters: { resource: 'text', operation: 'response' },
    position: [540, 300],
  },
  output: [{}],
});

export default workflow('mcp-test-min', '[mcp-test] msg model minimal')
  .add(webhookTrigger)
  .to(messageModel);

validate_workflow returns { valid: true, nodeCount: 2 }. create_workflow_from_code with the same code returns HTTP 500.

Reproduction 2 — full v2.3 OpenAI Message a model with credentials

Same shape with full parameters (modelId, responses.values, builtInTools, options) and newCredential('OpenAI'). Same outcome: validates, 500s on persist.

Reproduction 3 — AI Agent + OpenAI Chat Model (the canonical pattern)

import { workflow, trigger, node, languageModel, newCredential } from '@n8n/workflow-sdk';

const chatModel = languageModel({
  type: '@n8n/n8n-nodes-langchain.lmChatOpenAi',
  version: 1.3,
  config: {
    name: 'OpenAI Chat Model',
    parameters: { model: { __rl: true, mode: 'list', value: 'gpt-4.1-mini' } },
    credentials: { openAiApi: newCredential('OpenAI') },
    position: [540, 500],
  },
});

const aiAgent = node({
  type: '@n8n/n8n-nodes-langchain.agent',
  version: 3.1,
  config: {
    name: 'AI Agent',
    parameters: { promptType: 'define', text: '={{ \$json.body.prompt }}' },
    subnodes: { model: chatModel },
    position: [540, 300],
  },
  output: [{ output: 'agent response' }],
});

// + webhookTrigger and workflow().add(...).to(aiAgent)

Validates fine. create_workflow_from_code 500s.

Expected behavior

create_workflow_from_code should persist the workflow when validate_workflow accepts the same code. The validate/persist asymmetry forces a brittle workaround.

Actual behavior

create_workflow_from_code returns HTTP 500 with no error body for any workflow that includes a @n8n/n8n-nodes-langchain.* root node. Workflows using only n8n-nodes-base.* nodes persist successfully.

Impact

Anyone using the Builder MCP to scaffold workflows that include LLM operations is forced to:

1. Build the skeleton via MCP using placeholder Set nodes shaped to mimic the langchain output (e.g. `{ message: { content: 'TODO' } }` for v2.1, `{ content: 'TODO' }` for v2.3).
2. Open n8n in the browser.
3. Manually swap each placeholder for the real langchain node.
4. Bind credentials and update downstream references because the placeholder output shape rarely matches the real node exactly.

This adds 10-15 minutes of UI work per workflow and re-introduces the manual error surface the MCP is designed to remove.

Suggested triage

The asymmetry between `validate_workflow` (accepts) and `create_workflow_from_code` (500s) suggests the validation path doesn't exercise the same persistence code path. Either the validator is permissive about langchain nodes that the persistence layer rejects, or the persistence layer has a bug specific to the langchain namespace.

Happy to provide additional repros, MCP request/response logs, or test against any patched build.