Video generation MCP server: how Sume's generate_video works
Sume's hosted MCP server has a paid generate_video tool: a prompt or image in, a job id back, then jobs_wait and jobs_result for the clip.

A video generation MCP server gives an AI agent a tool that turns a prompt or an image into a video clip. Sume's hosted MCP server at https://mcp.sume.com/mcp has one, generate_video: it submits a paid video job, routes to sume/auto unless you name a catalog model, and returns a job id that the agent waits on with jobs_wait.
The contract comes from Sume's MCP overview, MCP tools and gates, Video Generation, and Jobs and results docs, read on 2026-09-27. The basics page says hosted MCP still works but is not the primary path today: it fits an agent that already speaks remote MCP, while a backend calls POST /v1/videos directly.
How do I connect an agent to it?
Point a remote MCP client at https://mcp.sume.com/mcp. With OAuth, mcp:read sessions see only read-only tools, so turn Write on at consent: mcp:write sessions see mutating and paid tools such as generate_video. An API-key session, with Authorization: Bearer $SUME_API_KEY, sees the full hosted tool set. Setup for Claude Code, Cursor, and Codex is in Connect Claude Code, Cursor, or Codex to Sume.
What does a generate_video call look like?
In current code, generate_video submits to POST /v1/videos, and every generation field goes inside payload, next to idempotency_key. This call only previews the cost; send it again with dry_run omitted or false to submit.
{
"idempotency_key": "desk-clip-001",
"dry_run": true,
"max_spend_usd": 2,
"payload": {
"prompt": "A vertical product clip on a desk, natural light",
"aspect_ratio": "9:16",
"duration": 5
}
}Which payload fields does generate_video take?
The fields follow Sume's POST /v1/videos contract. Read video-router_models before pinning a model, since limits differ per model.
| `payload` field | What it does |
|---|---|
prompt | Required. Text description of the video |
model | Omit to route to sume/auto, or send a catalog id from video-router_models |
duration, resolution, aspect_ratio | Each model advertises the values it accepts |
generate_audio | Whether to generate audio; defaults to the model's audio capability |
frame_images | First or last frame images, for image-to-video |
input_references | Reference images, for reference-to-video; frame_images wins if both are sent |
size, seed, provider.options | Rejected with a 400 (provider.options when non-empty) |
How does the agent get the finished clip?
It waits, then reads. In current code, generate_video answers with a job id within milliseconds, never a clip URL. Sume's docs say video generation typically takes 30 seconds to several minutes, depending on the model and parameters.
jobs_waiton the id holds at most 55 seconds per call (default 50), so the client's tool-call timeout must be longer than that.- On
wait_slice_expired, calljobs_waitagain with the same id. Never resubmit the paid create. - When the job is terminal,
jobs_resultreturns it. Sume returns generated outputs as Sume-hosted artifacts undermedia.sume.com. Waiting out long video jobs covers the loop.
Is Sume's generate_video tool free?
No. generate_video is a paid tool, billed from the workspace's USD balance and reserved on submit, and the wallet is the spend gate. There is no mcp:paid scope, so the controls are per call: idempotency_key is required, dry_run=true previews the cost without submitting the job, and max_spend_usd is enforced only when provided. GET /v1/videos/models lists each model with its pricing.
What doesn't it do?
- It does not name the model behind
sume/auto. Responses echosume/auto, and Sume does not disclose which family served the request. - It does not cut or assemble clips. Trim, filter, and Timeline are separate tools (video editing MCP server).
- It cannot read files from your laptop. Frame and reference images have to be reachable over public HTTPS.
Sources
Related posts
More in Agents
- What is Sume? A video agent platform, its API, and billing
Sume is a video agent platform: brief an agent in chat, save the recipe as a Format, and call it from your backend over one API. Surfaces and billing.
- YouTube chapter generator: timestamps from a transcript
A YouTube chapter generator turns a transcript into timestamps and titles from 00:00: sentence timings from speech-to-text, topic breaks from an LLM.
- Run the Sume video agent from your backend with Agent Completions
POST /v1/agent/completions runs the same agent as the Sume Agents chat, with tools and media generation, and returns an async run receipt you poll or webhook.
- Safe automation for AI agents that call paid APIs
Keep agents read-only by default, keep secrets out of logs, and on hosted MCP send an idempotency_key, preview with dry_run, and cap with max_spend_usd.
Written by Sume