Which video models can my agent call? video-router_models

Video model launches change weekly. An agent that hard-codes kling-3 or seedance-2.5 will miss new ids and keep using ones whose limits changed. Sume's MCP tools let the agent read the catalog first. The Video generation contract names two tools that matter here.

The two tools

The MCP payload schema is the same as REST: a drift test checks the keys and enums so MCP cannot accept something REST rejects.

An instruction that works

Tell the agent to list models once per session and to pick from what comes back. For example:

Before generating video, call video-router_models once.
Choose a model whose supported_durations contain the length I asked for
and whose supported_aspect_ratios contain my ratio.
Then call generate_video with that model id.
If none fit, say so; do not guess an id.

Why not leave it to sume/auto

The create tool name is generate_video, not video-router_create. The older name still resolves on call but no longer shows in the tool list.

• sume/auto takes 3 to 10 second clips in 16:9 or 9:16, with a default of 720p and 8 seconds, and Sume does not disclose which model served it. Anything outside that fails closed with 400 unsupported_capability.
• A named model gives you a length such as 30 seconds on seedance-2.5 or wan-3.0, and ratios such as 21:9 on Seedance.
• Never rely on a launch you read about: Kling 4.0, LTX-2.5 and Veo 3.1 are not catalog ids. The tool tells you the truth.