Claude MCP connector supports tool calls only: how Sume results arrive

Yes, the connector can drive Sume, because everything you need from Sume is reachable as a tool. Anthropic's page says that, of the MCP specification's feature set, only tool calls are currently supported. So a Haiku 5.5 agent submits a job with a paid tool, then reads status and results with jobs_status, jobs_result or jobs_wait, which are listed on Sume's hosted tool inventory.

What the connector skips

Do not design around a resource read or a prompt template; the connector will not do them.

Practical effect

Hosted MCP cannot read files from your laptop; the upload flow is create an upload URL, PUT the bytes from your client, then assets_complete. In a connector setup your server code does the PUT, not the model.

Keep each turn short. A job can take longer than a turn, so give the agent jobs_wait and let it return a job id if the wait runs out, rather than looping on the model.

A turn plan that fits the limits

Plan the agent's loop around jobs. Turn one calls the paid tool and gets a job id back. Turn two calls jobs_wait for that id. If the wait ends before the job does, the agent reports the id and stops, and your code starts a new turn later with jobs_status. When the job is complete, jobs_result returns the result.

Because only tool calls are supported, anything you want the model to see must come back as a tool result. Keep results small and ask for ids and URLs rather than media bytes.

• One job id per paid call.
• Never loop on a model turn to wait.
• Store the job id outside the conversation.