Bug Description

When using the n8n MCP SDK to create workflows via create_workflow_from_code (or validate_workflow), node-specific parameters placed at the top level of the config object are silently ignored by the serializer. The workflow is created successfully, all nodes appear on the canvas with their connections intact, and validate_workflow returns valid: true with the correct nodeCount. However, the resulting nodes have empty parameters, which only becomes visible when opening a node in the UI or attempting to execute the workflow. The root cause is that the serializer reads only config.parameters, while parameters placed directly under config (e.g. config: { mode, jsCode, ... }) are dropped. There is currently no warning or error to indicate that unknown keys at the top level of config were ignored. This is particularly problematic for LLM-based MCP clients (e.g. Claude via Claude.ai), which routinely skip the get_sdk_reference step when they pattern-match the SDK shape from prior context. The result is a "successful" workflow creation where everything looks correct until runtime, leading to a frustrating debugging loop because the first failure surface is a generic "Unknown error" at execution time.

To Reproduce

1. Connect a Claude.ai (or other LLM-based MCP client) to a self-hosted n8n instance via MCP.
2. Prompt the agent to create a minimal 3-node workflow: Manual Trigger → Set → Code.
3. The agent calls validate_workflow with the following SDK code (parameters at top level of config):

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

const manualTrigger = trigger({ type: "n8n-nodes-base.manualTrigger", version: 1, config: {}, });

const setFields = node({ type: "n8n-nodes-base.set", version: 3.4, config: { mode: "manual", assignments: { assignments: [ { id: "a1", name: "messaggio", value: "Test", type: "string" }, ], }, }, });

const codeNode = node({ type: "n8n-nodes-base.code", version: 2, config: { mode: "runOnceForAllItems", language: "javaScript", jsCode: "return [{ json: { ok: true } }];", }, });

const wf = workflow("TEST"); wf.add(manualTrigger.output(0).to(setFields)); wf.add(setFields.output(0).to(codeNode)); export default wf;

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

const manualTrigger = trigger({ type: "n8n-nodes-base.manualTrigger", version: 1, config: {}, });

const setFields = node({ type: "n8n-nodes-base.set", version: 3.4, config: { mode: "manual", assignments: { assignments: [ { id: "a1", name: "messaggio", value: "Test", type: "string" }, ], }, }, });

const codeNode = node({ type: "n8n-nodes-base.code", version: 2, config: { mode: "runOnceForAllItems", language: "javaScript", jsCode: "return [{ json: { ok: true } }];", }, });

const wf = workflow("TEST"); wf.add(manualTrigger.output(0).to(setFields)); wf.add(setFields.output(0).to(codeNode)); export default wf;

1. validate_workflow returns { valid: true, nodeCount: 3 }.
2. create_workflow_from_code creates the workflow successfully.
3. Open the workflow in the UI: nodes appear correctly connected on the canvas.
4. Open the Set node: the Mode and Assignments fields are empty.
5. Open the Code node: the Language, Mode and JS Code fields are empty.
6. Execute the workflow manually: fails with Problem in node 'Code' / Unknown error.

Expected behavior

validate_workflow should emit a warning (or, ideally, a hard error) when unknown keys are detected at the top level of a node's config. Something like:

WARNING: Node "Code" has unknown keys at the top level of config: ["mode", "language", "jsCode"]. Did you mean to place them under config.parameters?

This would short-circuit the silent failure: the LLM agent would see the warning during validation and self-correct before the workflow is created, instead of producing a "successful" empty workflow that only fails at runtime. Actual Behavior validate_workflow: returns { valid: true, nodeCount: 3 } with no warnings. create_workflow_from_code: returns a successful response with nodeCount: 3 and a working URL. Canvas: shows all 3 nodes correctly connected. Node parameters: empty. Execution: fails with generic Unknown error at the first node that needs the missing parameters. The correct syntax (which I have now confirmed works end-to-end on my instance) is:

node({ type: "n8n-nodes-base.code", version: 2, config: { parameters: { mode: "runOnceForAllItems", language: "javaScript", jsCode: "return [{ json: { ok: true } }];", }, }, })

Debug Info

MCP client behavior observed: search_nodes and get_node_types were called by the agent before writing the workflow code. get_sdk_reference was not called — the agent pattern-matched the SDK shape from prior context. This is the failure mode: agent skips reference fetch → guesses SDK structure → silent serialization drops unknown keys → validate_workflow returns green → user discovers the bug only at runtime. Reproducible workflow (post-fix on my instance): created with the incorrect syntax first (empty parameters confirmed), then updated with the correct syntax (config.parameters) and executed successfully (execution 94756, status: success, output: {"upper": "TEST FIX MCP RIUSCITO", "elaborato": true}). Suggested fix priority: I'd rank a validation warning as more impactful than docs improvements, because LLM agents routinely skip reference-fetch steps when they think they already know the API shape. A loud failure at validation time would catch this entire class of issues before workflow creation.

Operating System

Linux (Ubuntu, Hostinger VPS)

n8n Version

2.17.7

Node.js Version

v24.13.1

Database

SQLite (default)

Execution mode

main (default)

Hosting

self hosted