MCP avatar-image-to-video_create: choose Fabric or H3 Max

On Sume's hosted MCP server, avatar-image-to-video_create is the one paid tool for a talking clip made from a still and an audio file. Its model field selects veed/fabric-1.0, the default, or minimax/h3-max/lip-sync. Use generate_video for anything else; it refuses the lip-sync id with wrong_tool.

This post covers when an agent should choose each model and what the call looks like. It uses Sume's MCP tools and gates and models docs.

What does the tool take?

The same body as the HTTP routes, inside payload: audio_url, a measured duration_seconds, and exactly one visual source, either an image_url or an avatar_handle. The tool is listed under Avatars and talking-head in the hosted inventory and is paid, so like every paid tool it needs an idempotency_key. dry_run=true previews cost without submitting, and max_spend_usd caps a call when you provide it.

Sume's guidance for agents is to prefer the generated, inspected posed still as image_url, and to use avatar_handle only when the user named that avatar.

{
  "idempotency_key": "talking-clip-2026-10-03-001",
  "dry_run": true,
  "payload": {
    "model": "minimax/h3-max/lip-sync",
    "image_url": "https://example.com/host-still.png",
    "audio_url": "https://media.sume.com/example/line.mp3",
    "duration_seconds": 8.2,
    "resolution": "768p"
  }
}

When should the agent pick H3 Max?

Only when the audio fits its window, 5 to 14.8 seconds. Fabric stays the default for everything else, including lines shorter than 5 seconds and longer ones, because the image-to-video body accepts a duration_seconds of 1 to 300.

Sume refuses a duration outside the H3 Max window rather than clamping it, since a clamped reservation would render a clip shorter than the audio and drop speech. Never re-cut audio to fit a model; send the clip that needs Fabric to Fabric.

What should the agent do before and after the call?

Before: tools_schema for the tool name, then a dry_run and a look at balance and queue. After: poll with jobs_status or jobs_wait and read jobs_result. jobs_wait holds a single call for a limited time, so treat an expired slice as a reason to wait again, not to submit again.

In reports, use Sume public ids and media.sume.com URLs, and never paste signed URLs, OAuth tokens or API keys into chat logs. If a session has only OAuth mcp:read, mutating and paid tools are hidden and calls return insufficient_scope.

How do I write the agent instruction?

Give the agent the rule, not the model name. For example: measure the audio first; if it is 5 to 14.8 seconds, payload.model may be minimax/h3-max/lip-sync; otherwise omit model and Fabric is used. Always call tools_schema with name: "avatar-image-to-video_create" once per session to read the live contract, because the hosted catalog, not this post, is the source of truth.

Keep the Agents @ catalog in mind too: the lip sync model appears there as @minimax-h3-max-lip-sync, next to MiniMax H3 Max, so a user can name it explicitly. When they do, honor it, then check the window before submitting.