MCP server for video editing: which tools can an agent call?
An agent can trim, strip audio, filter and compose clips through Sume's remote MCP endpoint. Here are the tool names, order of calls, and what each step costs.
Connect Claude Code to https://mcp.sume.com/mcp and the agent gets a set of video editing tools: video_trim, audio_detach, video_filter, and the timeline_* family for assembling clips. Each one works on a clip that already lives on media.sume.com, so the first step is always an import. The edit tools never fetch from the open internet.
Connect it
The command in Claude Code's MCP docs adds an HTTP server by URL. Authenticate with /mcp, and turn Write on at the consent page, since cutting a clip creates a new file and is a write.
claude mcp add --transport http sume https://mcp.sume.com/mcpThe tools and their order
| Step | Tool | What it does | Rate in the docs |
|---|---|---|---|
| 1 | media-imports_create | Brings a clip into the workspace | Not listed here |
| 2 | video_trim | Cuts [start, end) or start plus duration into a new MP4 | $0.02 per job |
| 3 | audio_detach | Pulls the audio track into a new wav or mp3 | $0.01 per job |
| 4 | video_filter | Dim, crop or an allowlisted filtergraph; check_only: true is free | $0.02 per encode |
| 5 | timeline_create, timeline_compose, timeline_audio | Sequence clips and audio spine | Check GET /v1/catalog |
| 6 | video-captions_get | Reads captions | Not listed here |
How a call flows
Trim, detach and filter are asynchronous jobs. The pattern is the tool, then jobs_wait, then jobs_result. The result returns a new artf_ URL, never the source, and the trimmed MP4 is meant to go into timeline_create video[] with source_in 0. Every write needs an idempotency_key, so retries return the same job rather than a second charge.
The editing tools are deliberately narrow. Trim cuts one range from one clip, filter changes pixels on one clip, and detach pulls audio from one video. Anything that joins clips, sets transitions or lays an audio spine belongs to the timeline tools. Keeping that split in the agent's instructions avoids a common failure where a model tries to stitch a sequence with repeated trims and ends up paying for work the timeline would have done in one call.
Use the free check
video_filter accepts check_only: true, which runs the same checks as the encode without billing. Have the agent call it first and fix a bad filtergraph before the paid call. Under OAuth the paid call needs mcp:write. With an API key the full tool set is visible, so ask the agent to pass max_spend_usd or run dry_run before the first submit.
Two cautions
Rates are the public numbers in Sume's model pages, and they can change, so read the live price from the catalog before budgeting a batch. Also, audio_detach is meant for detaching once and splitting afterwards with timeline_audio, so do not detach the same clip repeatedly inside a loop.
For long unattended batch editing, an API key header avoids a sign-in prompt mid-sequence; Sume's MCP accepts OAuth access tokens or API keys.
Sources
More in Developers
- MiniMax H3 on Sume: 9 images, 3 videos, 3 audio, 12 in total check
On Sume, minimax-h3 and minimax-h3-max take at most 9 reference images, 3 videos and 3 audio clips, 12 in all, and audio cannot be alone. Check it in Python.
- Music-only Short: Timeline silence mode with a soundtrack bed
Build a Short with no voice: audio mode silence plus a soundtrack. The bed plays alone at the default -16 dB with no amix loss. Fields, limits, tail rule.
- Nearest supported aspect ratio per Sume image model, in Python
A model rejects an aspect ratio it does not list. Read the catalog, pick the closest ratio it accepts, generate, then crop to the exact shape you need.
- NestJS raw body controller to verify an AI video webhook signature
Enable rawBody in NestFactory, read req.rawBody in a controller and check Sume's sume-v1 HMAC before you act on a job.completed video event.
Written by Sume