Cut a long talk into short vertical clips with the Sume API

To turn one long recording into short clips with Sume, import it, run video-inspect with transcribe: true and sentence segmentation, choose ranges from the segments, cut each range with video-trim, then burn captions with video-captions. The source can be at most 1800 seconds (30 minutes) for inspect and trim. Choosing which sentences make a good clip is your step, or your agent's.

The four calls

Step by step

The whole flow is asynchronous after the import. Writes default to async for trim and captions, while inspect defaults to a sync wait of up to 30 seconds and returns a queued job if it takes longer. Treat every response as possibly a 202 and poll the job envelope, and you will not need two code paths.

• Inspect with frames: false if you only need the transcript, and segmentation: { mode: "sentence" } for caption-line shaped segments[] with no gaps. Send language_code as a hint, or leave it for auto-detect. A silent clip answers inspect_source_has_no_audio.
• Choose ranges from the segments, so a cut starts at a sentence start and ends at a sentence end.
• Trim with start plus either end or duration. precision: exact is the default and frame-accurate. keyframe is a stream copy that can start up to one group of pictures early, so re-base your times on actual_start_seconds.
• Send the trimmed video_url to captions. Each write needs an Idempotency-Key; poll the job at /v1/jobs/{id}/status.

What it adds up to

For a 30-minute talk cut into eight clips, the listed rates are 30 minutes of transcript at $0.01, eight trims at $0.02 and eight caption jobs at $0.20: $0.30 + $0.16 + $1.60 = $2.06. The inspect job also reserves its compute ceiling, so that figure is the sum of the per-unit rates and not a quote. Read the live prices from GET /v1/catalog.

The vertical part

video-trim keeps the source's frame unless you ask for an output of width, height and fps. If you need a vertical frame that reframes the shot, timeline-1.0 composes slots on video[] with fit set to cover, contain, stretch or blur, and renders 1080 by 1920 by default at $0.10 per started output minute. Run POST /v1/timeline-1.0/plan first; it is unbilled and returns estimated_cost_usd_micros.

The hosted MCP server exposes the same steps as video_trim and the other media tools, each followed by jobs_wait and jobs_result, which suits an agent that picks the ranges itself. Writes there need an idempotency_key, so build one from the source id and the range, and a retry of the same cut will not create a second file.