Make MCP Client: test Sume with Call a tool before AI runs it

Yes, Make can call Sume's tools through its MCP Client app: create a connection with Sume's hosted MCP URL, run the Call a tool module first to check the connection, and only then use Execute an action with AI. Make's own page recommends this order, and it matters more with Sume because some tools spend money.

This is the opposite direction from the Make-as-MCP-server posts on this site. Here a Make scenario is the client, and Sume is the server. Sume is not an official Make app: the Sume docs list no Make integration, so this is a generic MCP connection, not a one-click one.

What Make's MCP Client app documents

Everything in the table below comes from Make's MCP Client page. The page states no timeouts or limits for these modules, so this post does not invent any.

Connect Sume

Sume's production MCP endpoint is https://mcp.sume.com/mcp, per the MCP quickstart. Hosted MCP accepts OAuth tokens or Sume API keys, and the MCP auth page says an API-key session sends Authorization: Bearer or x-api-key. Make's page does not say which header its key field produces, so if the connection test returns 401, that is the first thing to check. We have not confirmed that Sume appears in Make's verified-server list, so use the manual URL path.

MCP Server: custom URL (manual)
URL: https://mcp.sume.com/mcp
Authentication: API key / access token
Value: your Sume API key (store it in the connection, not in a module field)

Test with Call a tool, read-only first

Add a Call a tool module and request tools_list. It is a read tool that lists everything the session can see. A second safe call is mcp_health, which confirms the endpoint and the auth source. Both are listed as useful first calls in the quickstart.

If you connect through OAuth instead, hosted Sume grants read-only access by default (mcp:read), and write or paid tools return insufficient_scope until the Write toggle is turned on at consent. That makes OAuth a safe choice for a scenario that should only look things up. Whether Make's OAuth flow completes against Sume's consent page is not something Make's page documents, so test it before relying on it.

Why the order matters for paid tools

Execute an action with AI lets a model choose tools. With an API key, Sume's hosted tool set includes paid tools such as generate_image, and writes and paid calls must include an idempotency_key. The optional dry_run preflights cost, and max_spend_usd is enforced only when you provide it, per the tools and gates page.

So give the AI module a task description that names the cap, and prefer Call a tool with a fixed tool name for anything that must run exactly once. Long renders should return a job id and be checked later with jobs_status, not re-submitted.

• Use Call a tool with tools_list to confirm the connection.
• Keep the first AI task read-only (list jobs, read a result).
• Pass an idempotency_key on every write or paid call.
• Never retry a paid create because a wait expired; read the job instead.

What to use instead when you need exact control

If a scenario must start one specific render and wait for a callback, the plain HTTP module with a webhook_url is simpler and is the documented path. See the jobs and results page for the async and webhook modes.