Claude MCP connector or Sume Agent Completions for a long video job?

Use Sume Agent Completions when the job is long and nobody is watching, and use Claude's MCP connector when a Claude conversation should drive Sume tools turn by turn. Agent Completions return 202 and a receipt that you poll, with a required spend cap. The MCP connector is a beta feature of the Claude Messages API that lets Claude call tools on a remote MCP server, so your code runs a conversation and carries the Claude side of the cost.

The two shapes

The two shapes differ in who runs the agent loop, as the table shows (read 2026-10-08).

Why long jobs favor the receipt

A render can outlast any single request. With the connector you manage the wait: call jobs_wait in slices up to 55 seconds, keep the conversation alive, and avoid a duplicate paid create when a request times out. With Agent Completions, the create call returns at once and you poll status_url until next_action is not poll_status. Statuses are queued, processing, completed, failed and canceled. You can send a communication.webhook_url to be told at a terminal status.

The connector fits when the model must reason between tool calls, for example choosing among catalog options a user is discussing. Anthropic's page also notes the connector is not ZDR-eligible and is not available on Bedrock or Google Cloud.

A minimal Agent Completion

The key must carry agent_completions:write to create and agent_completions:read to poll; an older key without them gets 403 insufficient_scope. Service-account keys cannot create completions.

curl -sS -X POST https://api.sume.com/v1/agent/completions \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Idempotency-Key: render-2026-10-08-001" \
  -H "Content-Type: application/json" \
  -d '{"instruction": "Make a 10 second product teaser.", "generation_spend_cap_usd": 5}'
# then poll
curl -sS https://api.sume.com/v1/agent-runs/$RUN_ID \
  -H "Authorization: Bearer $SUME_API_KEY"

Decision rules

The shortest rule is who owns the loop. If your code and your model should own the plan, use the connector or your own MCP client. If Sume's agent should own the plan and you want a receipt, use Agent Completions.

The second rule is attendance. A person in a chat can approve spend, so a connector-driven conversation can ask first. An unattended backend cannot, which is why Sume makes the cap mandatory on the completion route and why you should send max_spend_usd on every paid MCP call.

• Interactive and tool-by-tool: connector or MCP client with dry_run first.
• Unattended, one task per call: Agent Completions with a spend cap and webhook.
• Saved workflow with changing inputs: a Format run, not a completion.