Amazon Ads MCP Server (Lite) next to Sume's MCP server in one agent

Amazon announced an Amazon Ads MCP Server (Lite) at unBoxed 2026, and an agent could pair it with Sume's hosted MCP server so one tool makes the creative and the other handles the Amazon Ads side. What is not known yet is how you get the Amazon server: the announcement page does not say.

Treat this as a design note about two MCP servers in one client. It is not a setup guide for Amazon's server, which I could not confirm from its page.

A "Lite" label usually means a smaller tool set, but the page does not say that, so treat the name as a name and nothing more until Amazon documents it.

What did Amazon say about its MCP server?

The unBoxed page lists "Amazon Ads MCP Server (Lite)" as announced and says it simplifies agent interaction with core Amazon Ads tools. It gives no availability, no tool list, no authentication method and no pricing. The same page announces an Amazon Ads Agent with a conversational AI experience in expanded closed beta, and an AMC Expert Tool that generates SQL for analytics from natural language.

Until Amazon publishes a docs page, do not write instructions that assume tool names or scopes.

What does Sume's hosted MCP server offer?

Sume's remote server is at https://mcp.sume.com/mcp, and the quickstart shows adding it with claude mcp add --transport http sume https://mcp.sume.com/mcp. Under OAuth with only read scope you see read-only tools; paid and writing tools need mcp:write or an API key, and every write or paid tool needs an idempotency_key.

Waits are bounded. jobs_wait holds at most 55 seconds per call, so a long video render takes repeated waits on the same ids, never a second paid create.

claude mcp add --transport http sume https://mcp.sume.com/mcp

How would an agent use both servers?

Split the work by what each side owns. The table is a division of labor, not a recipe, since one half is unconfirmed.

What are the risks of two servers in one session?

Tool lists grow, and an agent has more ways to pick the wrong one. Use tools_list on the Sume side and read each tool's safety metadata before a paid call. Use dry_run and max_spend_usd for any burst of generations, because spend is wallet-based and there is no separate paid scope.

Keep credentials apart: the two servers authenticate separately, and nothing in Sume's docs says one can use the other's token.

Treat any agent run that mixes the two as two separate approvals. A prompt that says "make a video and attach it to my campaign" crosses an account boundary you may not want one instruction to cross, so keep the generation step and the campaign step as separate turns that you can review.

What should I do this week?

Nothing needs to be built yet. Confirm Sume's side works with your client, using the other MCP posts as a template, and watch Amazon's announcements page for an actual docs link. When it appears, check its tool names against tools_list before you write a prompt that mentions both.

When Amazon does publish docs, compare three things with what you already run: how the server authenticates, whether it can spend or only read, and whether it reports errors in a form an agent can act on. Those decide whether it is safe to leave in a shared client.