n8n instance-level MCP plus Sume MCP in one AI client

Yes, one AI client can talk to both: add n8n's instance-level MCP server and Sume's hosted MCP at https://mcp.sume.com/mcp as two separate server entries, each with its own credential. n8n's server exposes your workflows; Sume's server makes the media. Keep the tokens apart and the client can run a workflow that needs a clip, or ask Sume for a clip and hand the result to a workflow.

What does n8n's MCP server give the client?

According to n8n's MCP page, n8n uses an instance-level MCP server instead of per-workflow connections. You enable it under Settings, then Instance-level MCP, toggle MCP access, and copy a server URL and an access token into your client. The page lists Claude Desktop and Claude Code, ChatGPT, Google Gemini, Cursor, Codex CLI, Windsurf, Cline, Open WebUI and Google ADK as clients, and says it is available on all n8n plans from version 2.18.3.

Once connected, the client can search exposed workflows, create and edit them, trigger and test them, and read execution logs. A DEV Community write-up published October 6 describes the same expansion across agent hosts; it does not cite a dated n8n announcement, so treat the page above as the source.

What does Sume's MCP add?

Sume's hosted endpoint is https://mcp.sume.com/mcp. Per the Sume docs you authenticate in one of two ways that do not mix: OAuth, with a required mcp:read scope and an optional mcp:write scope, or an API key sent as Authorization: Bearer or x-api-key. Calls that change data or cost money must carry an idempotency_key, and an optional max_spend_usd caps the spend.

How do you keep the two credentials from crossing?

Give each server its own entry with its own header. Never reuse one token for both, and never paste either into a prompt. The Sume docs say not to store OAuth tokens in CLI config, paste them into prompts, or forward them to third-party providers, which also rules out putting a Sume token inside an n8n workflow that sends it elsewhere.

• Pick OAuth for Sume if you want a read-only consent first; write stays optional.
• Use an API key for unattended clients, sent in a header rather than the URL.
• Let n8n workflows call Sume through their own credential store, not through the chat client.

Which server should do the paid work?

Let Sume's MCP tools do the render, with a max_spend_usd on every paid call, and let n8n do the plumbing afterwards: filing the result, notifying a channel, updating a sheet. That way a retried workflow cannot silently re-buy a clip, because the paid step sits behind an idempotency_key you control. For a workflow that must call Sume itself, see the n8n MCP client post below.