How do I set up the n8n MCP server so Claude or ChatGPT can run my workflows?
Enable instance-level MCP in n8n, connect Claude or ChatGPT with OAuth or a token, expose only the workflows you choose, and fix the common connection problems.
- Time needed
- 30 min
- Steps
- 5
- Level
- Intermediate
In short: Turn on instance-level MCP in n8n (Settings > Instance-level MCP > Enable MCP access, as an owner or admin), copy the server URL, and connect Claude or ChatGPT with OAuth or a personal API token. Then expose only the workflows you want an assistant to run by marking them Available in MCP. Test with a harmless workflow first, because any client you connect sees every exposed workflow the connecting user can access.
Before you start
| Requirement | Detail |
|---|---|
| Role | Instance owner or admin to enable the feature; any user can connect a client to workflows they can access |
| URL | Your n8n instance must be reachable over HTTPS from the AI client |
| Reverse proxy | Must pass the MCP-Protocol-Version, Mcp-Method and Mcp-Name request headers |
| Version | Project and folder level exposure needs n8n 2.24.0 or later; the instance-level server is described in the current n8n docs |
| Kill switch | Setting N8N_DISABLED_MODULES=mcp removes the feature from a self-hosted instance |
This guide covers n8n's instance-level MCP server, which lets an assistant find and run your workflows. It is different from the MCP Server Trigger node, which turns one workflow into its own MCP server.
Step 1: Enable MCP access
- In n8n, open Settings > Instance-level MCP.
- Turn on Enable MCP access.
- Open Connection details and copy the Server URL. It ends in
/mcp-server/http.
Step 2: Choose how clients sign in
| Method | When to use it |
|---|---|
| OAuth (recommended) | Claude, ChatGPT and other clients that support OAuth sign-in. Each user approves access in a browser. |
| API key | Clients that only accept a bearer token. Use a personal MCP token from the Access token tab. Treat it like a password. |
If you use OAuth, restrict the allowed callback URLs in the OAuth client settings to the clients you actually use.
Step 3: Expose only the workflows you mean to
No workflow is exposed by default. Turn workflows on one at a time:
- Open the workflow's settings and enable Available in MCP, or use the workflow card menu, or the Workflows enabled list in the MCP settings.
- From n8n 2.24.0 you can expose a whole project or folder. Automatic exposure of new workflows is opt in; leave it off unless you have a reason.
- Prefer workflows with clear inputs and safe actions: "look up an order", "create a draft quote", "summarise yesterday's tickets".
Be aware of two behaviours from the n8n docs: search_workflows returns previews of every workflow the user can see, and access is not scoped per client, so every client connected as that user sees the same exposed workflows. execute_workflow runs the published version by default.
Step 4: Connect Claude or ChatGPT
- In the AI client, add a custom connector or MCP server and paste the Server URL.
- Sign in with OAuth when prompted, or paste the API key if the client asks for a bearer token.
- Ask the assistant to list available workflows, then run a harmless one with a test input.
Step 5: Add guardrails before real use
- Put approval steps inside workflows that send messages or change records. n8n supports human review for AI tool calls, and a Slack or email approval step works for any workflow.
- Use a dedicated n8n user with access to only the projects an assistant needs.
- Log inputs and results so you can review what the assistant ran.
Troubleshooting
| Symptom | Likely cause | Fix |
|---|---|---|
| Client cannot connect | Proxy strips MCP headers or the URL is not HTTPS | Pass MCP-Protocol-Version, Mcp-Method, Mcp-Name; check TLS |
| No workflows listed | None marked Available in MCP | Enable them per workflow, project or folder |
| OAuth fails | Callback URL not allowed | Add the client's callback URL to the allowed list |
| Wrong version runs | execute_workflow uses the published version |
Publish the workflow after edits |
| Feature missing | N8N_DISABLED_MODULES includes mcp, or not owner/admin |
Remove the setting or ask an admin |
FAQ
Do I need the MCP Server Trigger node?
Not for this. The instance-level server exposes chosen workflows to assistants. The MCP Server Trigger node is for building a single workflow as its own MCP server with custom tools.
Can I limit one client to certain workflows?
Not per client. Access follows the connecting user, so use a dedicated user with limited projects.
Is OAuth or an API key safer?
n8n recommends OAuth. If you use an API key, store it in the client's secret settings and rotate it if it leaks.
Does this work on n8n Cloud and self-hosted?
The n8n docs describe the feature for both; self-hosted instances can also disable it with N8N_DISABLED_MODULES=mcp.
How Octacer helps
Octacer sets up and secures n8n, including MCP access, approvals and logging. Related reading: the OpenClaw, n8n and MCP stack, MCP business ops starter pack, MCP security best practices and our automation services.
Next step
Pick one read-only workflow, expose it, and run it from your assistant this week. If you want help choosing what to expose, book a call.
Sources
- n8n Docs, Accessing n8n MCP server: https://docs.n8n.io/advanced-ai/accessing-n8n-mcp-server/
- n8n Docs, Human-in-the-loop for tools: https://docs.n8n.io/advanced-ai/human-in-the-loop-tools/
- n8n releases on GitHub: https://github.com/n8n-io/n8n/releases
Keep going
More guides
Want it running on your stack?
We implement this guide for clients, tailored to your stack, your data, and your team's workflow, and handed over documented.