Claude API MCP connector: public server, tool calls only, one toolset

The Claude API MCP connector can reach https://mcp.sume.com/mcp because that server is public and speaks HTTP, and it will use only the server's tools. It does not support local STDIO servers, and it supports tool calls only, so anything your agent needs from Sume has to be a tool. Add the beta header, one mcp_servers entry and one matching mcp_toolset entry.

What the request needs

Anthropic's connector page lists the shape of a request. You pass the beta header mcp-client-2025-11-20, since the older mcp-client-2025-04-04 header is deprecated. You add an mcp_servers array with the server's url, a name and an optional authorization_token, and you add an mcp_toolset entry to tools that names the server through mcp_server_name. Every server must be referenced by exactly one toolset, or the request is refused.

Limits to read first

Two limits matter for a video agent. The server must be reachable over public HTTP, in Streamable HTTP or SSE form, so a Sume setup that depends on a local process cannot work here. And the feature is in beta and is not eligible for zero data retention, which is a point for your own compliance review before you send anything sensitive through it.

Auth is your job here

The connector does not run an OAuth consent flow for you. It forwards the token you give it. For an unattended job, an API key is the simple option, and it gives the full tool set, so keep paid calls gated with idempotency_key, dry_run and max_spend_usd. If you want a read-only agent, use an OAuth session that only has mcp:read, and obtain that token outside the connector.

A request that checks the connection

The script below sends one message that asks Claude to call mcp_health on the Sume server and prints the content block types it got back. It needs ANTHROPIC_API_KEY and SUME_API_KEY. It only reads, so it spends nothing on Sume.

import json, os, urllib.request
body = {
    "model": "claude-sonnet-5-5",
    "max_tokens": 512,
    "messages": [{"role": "user", "content": "Call mcp_health on the sume server."}],
    "mcp_servers": [{
        "type": "url",
        "url": "https://mcp.sume.com/mcp",
        "name": "sume",
        "authorization_token": os.environ["SUME_API_KEY"],
    }],
    "tools": [{"type": "mcp_toolset", "mcp_server_name": "sume"}],
}
req = urllib.request.Request(
    "https://api.anthropic.com/v1/messages",
    data=json.dumps(body).encode(),
    headers={
        "x-api-key": os.environ["ANTHROPIC_API_KEY"],
        "anthropic-version": "2023-06-01",
        "anthropic-beta": "mcp-client-2025-11-20",
        "content-type": "application/json",
    },
)
with urllib.request.urlopen(req) as r:
    print([b["type"] for b in json.load(r)["content"]])

Reading the output

If the list shows mcp_tool_use and mcp_tool_result blocks, the connector reached the server and the tool ran. If the call fails, check the header value first, then the toolset name, then the key.

Long jobs in a short turn

For long video jobs, design the loop around jobs_wait, which holds for 50 seconds by default and no more than 55, and returns wait_slice_expired when the job is still running. A model turn that sees that code should wait again on the same job id, and never call the paid create tool a second time.