Field guide · n8n MCP authentication
“Authentication failed” is a headline. Capture the failing boundary.
Replacing the token again will not fix a wrong endpoint family, transport mismatch, rewritten header, or proxy path. First prove which n8n MCP surface you are calling and the last protocol step that succeeds.
Redacted support packet
Record facts without recording the credential.
The plain-text worksheet captures endpoint shape, transport, version, response evidence, and the smallest direct-client result.
1. Name the n8n MCP surface
A workflow-level server with separate test and production URLs. It supports Bearer or Header authentication and connects only to tool nodes attached to that trigger.
A server enabled under Settings with OAuth2 or a personal MCP access token. Individual workflows must be enabled for MCP access.
The n8n MCP Client Tool is the client. Its endpoint, transport, and credential configuration must match the external server—not either n8n server surface.
Write down the exact endpoint shape and which of these three roles n8n is playing. Do not paste the token into the worksheet.
2. Capture the real response
Record the exact status, response body, authentication challenge, final URL after redirects, timestamp, and request or correlation ID. A client may collapse several different 401 and 403 responses into the same “Authentication failed” label.
Run the credential-free endpoint check first to expose redirects, response types, initialize behavior, and OAuth protected-resource metadata without placing a token in the report.
If OAuth returns invalid_target, use the
local resource-identity check
to compare the endpoint, requested resource, and advertised
protected-resource metadata without uploading the URLs.
3. Match URL and transport before changing auth
- Use the production URL when the MCP Server Trigger workflow is published.
- Use the test URL only while n8n is listening for a test event or executing the unpublished workflow.
- Confirm the client transport matches SSE or Streamable HTTP for the URL you selected.
- Call the canonical final URL directly; record any host or path change in the redirect chain.
- For instance-level MCP, copy the server URL from the Connection details panel rather than reconstructing it.
4. Remove the agent from the test
Use a direct MCP client to connect, initialize, list tools, and make one low-risk tool call. The official MCP Inspector is designed for connection, capability, schema, and tool testing. Keep the credential in the client configuration; never include it in a screenshot, support post, or intake.
5. Verify the credential contract
For an MCP Server Trigger, confirm the node and client use the same documented Bearer or Header authentication contract. For instance-level MCP, distinguish the OAuth2 flow from the personal access-token flow. Compare header name, scheme, and non-secret length or fingerprint—not the raw value.
If the authorized direct client still returns 403, capture a redacted request-header name list at the final server boundary. That separates a header lost or rewritten in transit from a credential rejected after arrival.
6. Check the self-hosted request path
n8n documents that SSE and Streamable HTTP need persistent requests to
reach the same server instance. With multiple webhook replicas, route
all /mcp* traffic to one dedicated webhook replica. Behind
a reverse proxy, preserve the MCP path and disable buffering for that
endpoint.
7. Build a support-quality minimum reproduction
Use one new, isolated workflow or one instance-level connection and preserve the first failure. Include the n8n version, Cloud or self-hosted deployment, MCP surface, endpoint shape with identifiers redacted, transport, authentication type, exact status and body, timestamp, request ID, and direct-client result.
State the target outcome in one sentence: “initialize and tools/list succeed, then this one low-risk tool returns the expected result.”
Official references
- n8n MCP Server Trigger documentation ↗
- n8n instance-level MCP connection documentation ↗
- Model Context Protocol Inspector ↗
Agent Rescue Desk is independent and is not affiliated with, endorsed by, or sponsored by n8n or the Model Context Protocol project.
Still blocked after the direct test?
Submit the redacted boundary evidence.
The 12-Hour Agent Rescue covers one bounded n8n or MCP workflow, including the directly required authentication, proxy, runtime, and tool layers.