Sume hosted MCP audio tools: music, TTS, STT, detach, timeline

The hosted Sume MCP exposes music_create, tts_create and stt_create for generating audio, audio_detach for pulling audio out of a hosted clip, and timeline_audio for joining or splitting tracks. Two read tools, tts_source_get and tts_source_verify_spine, check a voiceover against an accepted script.

The names come from Sume's MCP tools and gates page and the MCP overview (checked 2026-10-10). Prices are those of the matching REST routes.

Which tool makes which audio?

Paid generation tools take an idempotency_key, and the docs say to omit payload.model so the call routes to sume/auto unless the user named a model family. That means a bare music_create goes through the router rather than a pinned engine.

What do the read tools check?

tts_source_get returns the accepted-script manifest that a tts_create call with transcript_source uses. tts_source_verify_spine compares the selected TTS jobs with that script, so an agent can confirm that the voiceover it has is the voiceover that was approved before it moves on to a render.

• Use transcript_source when a script was reviewed and accepted.
• Run the verify tool before placing audio on a timeline.
• Both read tools are free.

Which tools handle video-side audio?

video_inspect is the default for "what is in this clip", and audio_detach is the step before stt_create when you want a transcript from a hosted video. Detach returns a wav by default, or mp3 at 128 kbps, with channel and sample-rate options. Captions are read through video-captions_get; the older create tools are unlisted in tools_list, though in-flight clients can still call them by name.

What the MCP cannot do

Hosted MCP cannot read files from your computer. Upload goes through assets_upload_url, a client PUT of the bytes and then assets_complete. Voice cloning is not available as a tool, because it happens in the Sume app. If you need a pinned Sonic engine for repeatable narration, call the TTS Router over REST, where model is required.

A short sequence an agent can follow

A common order for a voiceover project is: tts_source_get to read the approved script, tts_create with transcript_source once per accepted section, tts_source_verify_spine to confirm the jobs match, then timeline_audio to join the parts. If the project starts from a recorded clip, the order is audio_detach, then stt_create on the extracted audio, and edits to the text from there.

Each paid call carries an idempotency_key, so a retried call does not make a second charge. Reuse the same key for a retry of the same step and a new key for a new step.

• Read tools are free and safe to repeat.
• Paid tools need an idempotency_key.
• Wait for jobs with jobs_wait before reading results.

Cost of that sequence

For a script of 6,000 characters in three sections, TTS is three jobs of 2,000 characters at $0.0475 per 1,000, which is $0.095 each and rounds up to $0.10 per job, so $0.30. The join is $0.01. Detaching a clip and transcribing one minute of speech adds $0.01 plus $0.01. These are the REST prices; the MCP uses the same routes.