I’m trying to execute an n8n workflow via the MCP integration from Claude (claude.ai).
The workflow is marked as “Available in MCP” and Claude can successfully:
-
Search for workflows (
search_workflowsworks) -
Get workflow details (
get_workflow_detailsworks)
However, when Claude attempts to execute the workflow using execute_workflow, the following happens:
-
Claude receives an immediate error:
"Unknown execution mode" -
An execution IS created in n8n and appears in the queue as “Starting soon”
-
The execution remains stuck in queue indefinitely
-
No input data is passed to the workflow — the trigger node shows no input/message
Workflow Setup:
The workflow has multiple triggers connected to an AI Agent node:
-
Chat Trigger (primary)
-
Webhook Trigger (POST)
-
Form Trigger
All triggers are connected to the same AI Agent that uses an MCP Client to call Apify scrapers.
What I’ve Tried:
-
Executing with
type: "chat"andchatInputparameter -
Executing with
type: "form"andformDataparameter -
Executing with
type: "webhook"andwebhookDataparameter -
Switching webhook from GET to POST
-
Adding a Form Trigger with defined fields
All attempts result in the same behavior: “Unknown execution mode” error returned to Claude, execution queued in n8n with no input data.
Expected Behavior:
The workflow should execute with the provided input data (e.g., a TikTok URL to scrape) and return the result to Claude.
Execution Attempts:
Claude’s MCP tool attempts with various input formats all return the same error:
{"success":false,"executionId":null,"error":"Unknown execution mode"}
Despite the error, executions appear in n8n’s queue — but with no input data, causing them to hang.
Questions:
-
Is there a specific input format required for MCP execution via Claude Desktop Connectors? In the release video, it said form, chat, and webhook should all work.
-
The video mentions “native connection into all of the most popular chat apps” is coming — is Claude Desktop Connectors the right method currently, or should we wait for native integration?
-
Why would executions queue with no input data when the MCP call includes input parameters?
-
Is there documentation on the expected payload format for
execute_workflow?