n8n MCP Client Tool with Sume: setup and the SSE caveat

You can point an n8n MCP Client Tool at Sume's hosted endpoint, https://mcp.sume.com/mcp, with a Bearer Auth credential that holds your Sume API key, but neither side documents that the pairing connects: n8n's docs call the node's URL field SSE Endpoint, while Sume documents its endpoint for streamable HTTP clients. Attach the node to an AI Agent, set Tools to Include to Selected, and have the agent call Sume's mcp_health tool before you build on the connection.

n8n's side comes from its MCP Client Tool, Tools Agent, and Human-in-the-loop for tools pages, among others listed under Sources; Sume's side comes from OAuth and API keys, MCP quickstart, MCP tools and gates, and Jobs and results, all read on 2026-09-27. Sume has no n8n node; the MCP Client Tool would talk to Sume's remote MCP server directly. Sume's basics page says hosted MCP still works but is not part of the primary path today. If the connection fails, or the workflow needs no agent, n8n AI video workflow: resume a Wait node on a Sume webhook calls Sume's HTTP API instead.

How do I set up the MCP Client Tool for Sume?

An AI Agent node needs at least one tool sub-node; attach the MCP Client Tool to its Tool connector. n8n stores credentials securely and tests each one when you save it. Then set three fields:

• Endpoint: https://mcp.sume.com/mcp, in the field n8n's page calls SSE Endpoint. n8n lists a Server Transport setting for its standalone MCP Client node, which runs MCP tools as regular workflow steps, but not on the tool node's page, so check which transport options your node shows.
• Authentication: Bearer Auth, with your Sume API key as the token. n8n's Bearer auth is header auth named Authorization with the value Bearer <token>, which is how Sume takes a key; a generic Header Auth credential named x-api-key also fits.
• Tools to Include: Selected, then pick the tools. All Except instead exposes every tool you don't pick.

Which Sume tools should the agent see?

An API-key session sees the full hosted tool set, write and paid tools included, so Tools to Include is your first filter. n8n's release notes say MCP Client Tool nodes prefix each tool name the agent sees with the sanitized node name, as in MCP_Client_get_weather, while the server still receives the original name. n8n's guide for another MCP server warns that a system message naming the short tool names makes the model call names that don't match, so describe the steps instead of naming tools.

How do I require approval before a paid Sume call?

Use n8n's human review for tools. Click the tool connector on the AI Agent node, find the Human review section, pick an approval channel such as Chat, Slack, or Telegram, and connect the tools that need approval to the review step. When the agent calls one, the workflow pauses and the reviewer sees the tool and its parameters; approve and it runs with the agent's input, deny and it is canceled and the agent is told.

Review applies to the tool nodes you connect to it, so split Sume across two MCP Client Tool nodes: one with the read tools on the agent, one with only the paid tools behind the review step. n8n says to describe the review setup in the system message. The reviewer should check Sume's gates in the parameters: an idempotency_key on every paid call, dry_run=true for a preview that submits nothing, and max_spend_usd when you want a cap.

How long can the agent wait for a render?

One jobs_wait call holds at most 55 seconds, and after wait_slice_expired the agent should wait again on the same ids and never resubmit the paid create; MCP tool call timeouts on long-running video jobs covers the rest. Waiting also spends agent iterations: in n8n's guide for another asynchronous MCP server, each poll costs one iteration, the tool call plus the model turn that decides to poll again. The Tools Agent's Max Iterations option defaults to 10, so raise it for renders that need several waits.

Stopping early cancels nothing on Sume's side: a client-side timeout does not cancel the job, which keeps running and still bills. For renders that take minutes, the Wait-node pattern in the post linked above avoids holding an agent loop open.