ElevenLabs Music Skill

Make songs, instrumentals, jingles, loops and video scores with the elevenlabs_music_* tools, edit or extend songs they made, and split any song into stems. Every generation is billed to the user's ElevenLabs account by length; drafting a plan and listing finetunes are free. Sound effects (one-shot foley, UI blips) belong to the ElevenLabs Sound Effects plugin, not here.

When to Use

• The user wants music: a song with vocals, an instrumental, a jingle, background music, a loop, a theme.
• They give lyrics or want exact structure (verse/chorus/bridge, timings).
• They want to change part of a song made earlier, make it longer, add an intro/outro, loop it, or make "another one like this".
• They want the vocals or drums of a song on their own (stems), or music for a video.

Prerequisites

• The plugin's tools are listed (they appear only once ELEVENLABS_API_KEY is set). The Music API needs a paid ElevenLabs plan; a 401/403 names the fix.
• Local files the user mentions (audio to split or upload, videos to score) must be on the machine Hermes runs on; pass absolute paths, or paths relative to the working directory.

How to Run

Tool	Billed	Use for
elevenlabs_music_compose	yes	A song or instrumental from one prompt (3 s to 10 min)
elevenlabs_music_draft_plan	free	ElevenLabs writes a composition plan (sections, lyrics, styles, durations) from a prompt
elevenlabs_music_compose_plan	yes	Generate from a plan: exact lyrics/structure, and every edit, extension, loop or "sounds like"
elevenlabs_music_upload	yes	Store the user's own song at ElevenLabs so plans can edit or sample it
elevenlabs_music_stems	yes	Split a track into 2 stems (vocals, accompaniment) or 6 (vocals, drums, bass, guitar, piano, other)
elevenlabs_music_from_video	yes	Music timed to 1-10 videos, optionally with sound effects
elevenlabs_music_finetunes	free	Finetune ids (own, workspace, ElevenLabs-curated styles such as reggaeton or 80s nu-disco)
elevenlabs_music_job	free	Check or wait for a generation that is still running

Quick Reference

• Results carry media_tag (or media_tags), track_id, title, lyrics, sections (label + ms range of each part), composition_plan and elevenlabs_song_id.
• elevenlabs_song_id + a range is how a plan keeps part of an earlier song. It works only while the result says stored_for_editing: true.
• Plan limits: 1-30 chunks, each 3-120 s, total 3 s to 10 min; at most 30 lyric lines of 200 characters per chunk; up to 50 styles per list; a conditioning_ref covers at most 30 s.
• Models: music_v2_5 (default, newest), music_v2, music_v1. Plans need v2 or v2.5.

Procedure

1. Write the prompt (from ElevenLabs' prompting guide)

A prompt answers five questions: genre, mood, instrumentation, tempo (BPM), production era. Any left open gets the most average answer. Add a key ("in A minor") for layering. Studio words move real levers: sidechained, close-mic'd, bone-dry, tape saturation, plate reverb. Narrate arrangement in order ("start with just a shuffled drum loop, add a sub bass after four bars, then chopped vocal stabs"); just and then matter. Vocal delivery: whispered, belted, breathy, deadpan, stacked harmonies, "two singers harmonizing". Lyrics can go straight into the prompt. Loops are about exclusion: bars, BPM, key, and what is banned ("four bars, no melody, just drums").

• A length the user names goes in length_seconds, not only in the prompt.
• Instrumental: instrumental: true (it guarantees no vocals).
• Many languages work; say which ("Spanish lyrics").

2. Pick the route

• Quick idea, no exact lyrics: elevenlabs_music_compose.
• User's lyrics or exact structure: elevenlabs_music_draft_plan (free) → show the sections and lyrics → apply the user's edits to chunks → elevenlabs_music_compose_plan with those chunks and the same model. Put the user's lyric lines in each chunk's text after a [Section] name; overall vocal style goes in positive_styles, short cues in braces ({guitar solo}), sung sounds in parentheses ((ooh)). The first chunk's styles set the song: give it 6+.
• A style the user named that matches a curated finetune: elevenlabs_music_finetunes with created_by: elevenlabs, then pass its finetune_id and its model_id as model.

3. Edit, extend, loop, re-style (plan chunks over a stored song)

Read the earlier result's sections, length_seconds and elevenlabs_song_id, then build chunks:

Goal	Chunks
Redo one section	slice 0..start + new chunk for that section (its old text/styles, changed as asked) + slice end..length
Extend / add an outro	slice 0..length + new chunk ([Outro], styles from the first chunk)
New intro	new [Intro] chunk + slice 0..length
Seamless loop	slice of the phrase + short glue chunk (context_adherence: high) + the same slice again
"Another song like this"	new chunks, first one with conditioning_ref = a ≤30 s slice + condition_strength

A slice is {"song_id": "<elevenlabs_song_id>", "range": {"start_ms": a, "end_ms": b}}. Billing counts the whole result's length, kept slices included. For the user's own file: elevenlabs_music_upload first (billed like a generation of that length), then use its elevenlabs_song_id and sections.

4. Deliver

• Every surface: put each media_tag line from the result on its own line in the reply, copied verbatim. Add the title and a one-line description; offer next steps (variations, extend, stems).
• Hermes Desktop: the line becomes a music card with synced lyrics, download per format and a Next menu; don't describe buttons the user can see.
• Chat apps (Telegram, Discord, Slack, WhatsApp, Signal, ...): a generation that is not done within a few seconds returns status: running with a job_id and a next_step. Follow it exactly, in the same turn: first arm terminal(command="sleep <N> && echo elevenlabs music job <job_id>", background=true, notify=true) with no message before it, then reply in one line that it started (with the estimate) and end the turn. A foreground sleep holds the chat just like waiting. When the wake arrives, call elevenlabs_music_job with wait_seconds: 60 and send the media_tag lines. If the user asks "is it done?" meanwhile, check with wait_seconds: 0, deliver if finished, and answer the later wake with "already sent above". If the terminal tool or its notifications are unavailable (cron, hermes chat -q), call elevenlabs_music_job with wait_seconds: 240 instead. Cards can't be clicked there: end with next steps as a short numbered list.
• Terminal: the result's file_path is the file; say where it is.

Pitfalls

• Never start the same request again because it is slow or a job is still running: it bills twice. Check the job instead.
• elevenlabs_music_compose_plan refuses music_v1; drafts and edits use v2 or v2.5.
• An edit needs stored_for_editing: true on the source result; otherwise upload the file first.
• Section names in brackets alone ([Outro]) make an instrumental part; lyric lines make it sung.
• Copyrighted lyrics, artist names or uploads can be refused by ElevenLabs; say so and offer an original.
• Files go to <HERMES_HOME>/cache/generated/audio/; copy one into the user's project when asked.

Verification

• The reply contains each media_tag line exactly as returned (one per file for stems and video scores).
• For a running job in a chat app: a one-line "started" reply plus exactly one background wake, and the delivery turn sends the files once.
• For edits: the result's sections show the kept slices where the user expected them.