How an agent picks a video model over hosted MCP

Over hosted MCP, an agent chooses a video model in two calls. It reads the catalog with video-router_models, picks an id from the result, then calls generate_video with payload.model set to that id and a required idempotency_key. If payload.model is omitted, the call routes to sume/auto. The finished clip is fetched with the jobs tools.

Facts from Sume's hosted MCP docs and tools and gates, read 2026-10-01, plus the Video Generation docs for model limits.

What are the steps?

The tool names below are the underscore ids in tools_list; the docs say that list is authoritative for your session, so ask for it first rather than trusting a post.

How should the agent choose?

Give the agent a rule rather than a free choice. If the brief names a duration, filter the list to models whose supported_durations include it: for example 3 to 10 seconds for gemini-omni-flash-1.1, 4 to 30 for seedance-2.5, 2 to 30 for wan-3.0, and at most 15 for most others. If the brief has a source video or reference photos, filter on supported_input_references. If the brief says cheapest, ask for the cost first. If it does not matter, omit the model and accept sume/auto, remembering that the response will not say which family ran.

Use a stable idempotency_key per shot, such as the shot id, so a retried tool call after a timeout does not create a second paid job. The docs say the key is for transport and deduplication, not human approval; the wallet and admission checks are the spend gate.

Limits

Not every product is on hosted MCP: Sume Video 1.0 (videos_create) is REST-only, and the Video Router tools are the MCP path. OAuth clients need mcp:write for paid calls. I have not described the exact argument schema beyond payload.model and idempotency_key, so call tools_schema for generate_video before building the call.