Your agent calls generate_video, or Sume's agent does: two meters
With hosted MCP your model client runs the loop and you pay its provider; with an Agent Completion Sume's agent runs it and you set generation_spend_cap_usd.
There are two ways an AI agent can produce video through Sume, and they put the loop in different places. With hosted MCP, your own client and model, say Claude Haiku 5.5, drive the loop and call tools like generate_video. With an Agent Completion, Sume's agent runs the loop and you only send the task and a generation_spend_cap_usd. Either way generation is charged on Sume's rates; the difference is who chooses the model, who holds the tool loop, and where the cap lives.
The docs say an Agent Completion accepts only model: sume-agent, so you cannot select Haiku 5.5, Mistral Large 4 or any other model for it. If you need a specific model reasoning over the tools, use the MCP route with a client that runs that model.
Side by side
From the MCP, Agent Completions and Safe automation pages, read 2026-10-09.
| Question | Hosted MCP from your client | Agent Completion |
|---|---|---|
| Who runs the model loop | Your client and chosen model | Sume's agent (sume-agent) |
| Entry point | https://mcp.sume.com/mcp | POST /v1/agent/completions |
| Auth | OAuth or API key | API key with agent_completions:write |
| Spend control | Optional max_spend_usd, dry_run, wallet admission | Required generation_spend_cap_usd |
| Result | Tool results in your client | Async receipt; poll status_url or use a webhook |
| Needs a person present | Interactive OAuth, or a key for automation | No person has to monitor it |
Choosing
Pick hosted MCP when the model matters to you: a cheap small model for routing, a stronger one for planning, or a client your team already uses such as Cursor or Claude Code. Pick an Agent Completion when your backend only has a task and you want Sume to carry out the steps with a hard cap on the generation spend.
Neither is the Studio Agent product, which is not a public MCP connector. Do not point a client at it.
What each meter covers
The MCP route has two meters. The model provider bills tokens at its own prices, for example Haiku 5.5 at $0.10 input and $0.50 output per million tokens for short prompts, and Sume bills generation at its catalog rates. The Agent Completion's documented cap covers generation spend, and the page does not describe a separate token charge, so check the pricing page before you assume anything about it. For a cap, start from the catalog arithmetic: three 5-second Wan 3.0 720p clips are 3 x 5 x $0.125 = $1.875, so a cap of $2.00 covers them with a small margin.
A checklist for either path
Whichever route you take, treat spend as something you configure, not something you hope for.
- MCP: require
idempotency_keyandmax_spend_usdin the instructions, and preview withdry_runbefore a burst. - Agent Completion: pick the cap from the catalog price of the clips you expect, plus a margin, and send
Idempotency-Keyon the request. - Both: read the result from durable
media.sume.comURLs, and keep signed URLs and keys out of logs.
Sources
Related posts
More in Agents
- Can I create a Sume schedule over the API? Not today
Sume schedules are created in the dashboard with a cron and timezone or an API-only trigger. The API cannot create or edit them, but it can fire and read runs.
- Copilot Studio hooks fail open: cap paid media spend server-side
Copilot Studio hooks can block a tool call, but a failed hook lets the agent continue. Put the media spend limit where it can't fail open: the Sume run cap.
- Gemini agent keeps memory for days; Sume API runs start fresh
Google's Gemini agent keeps one memory across devices and days. Sume's API runs start in a new thread, so state travels as input, files and previous_run_id.
- Guardrails for an AI agent that calls paid media APIs
Five guardrails for an agent that spends money: required cap, read before paying, idempotency keys, no secrets in logs, and a cancel path. Drawn from Sume docs.
Written by Sume