creativeclaw-generate-audio

Generate Audio

Read shared execution guidance once per task before using tools. It covers existing authorization, model discovery, optional cost checks, imports, and recovery.

Create a finished sound effect, ambience bed, Foley cue, or music track with ElevenLabs through Creative Claw. Keep speech in creativeclaw-generate-voiceover; generate_speech and generate_audio use different provider endpoints and parameter contracts.

Route the request

Requested result	Tool and model
Narration, dialogue, or a speaking character	Use creativeclaw-generate-voiceover and generate_speech
Sound effect, Foley, transition, impact, texture, or ambience	generate_audio with sfx/elevenlabs-sound-v2
Music, score, bed, sting, jingle, or song	generate_audio with music/elevenlabs-music-v1
Add one finished audio track to video	merge_media with operation: "merge_audio_video"
Join audio clips one after another	merge_media with operation: "merge_audios"

Do not send a sound-effect or music model to generate_speech. Do not describe concatenation as mixing: merge_audios joins clips sequentially and does not layer narration, music, and effects.

Workflow

1. Identify the result as speech, sound effect, ambience, or music. Clarify duration and whether the audio must loop, end cleanly, contain vocals, or match picture.
2. Choose the SFX or music model from the requested outcome and fetch missing settings with get_model_params. Use list_models({ category: "audio" }) when discovery is needed.
3. When the user wants inspiration or a starting point, use creativeclaw-find-examples with output_type: "audio" and load the chosen example. Otherwise write the production prompt directly. Read references/prompting.md for the relevant pattern.
4. Use estimate_generation with operation: "audio" only when the user asks about cost/balance or gives a budget constraint. State consequential settings without a routine approval question.
5. Call generate_audio. Keep each requested variation as a separate generation so the user can audition it independently.
6. Listen for timing, unwanted voices, clipping, noise, loop seams, weak endings, and whether the sound matches the described space and intensity. Revise the prompt based on the audible defect.
7. Preserve the approved permanent audio URL. Name and tag the asset when useful for a larger project.
8. Read media-assembly.md before combining segments or adding audio to video. Check durations, follow merge continuations, and distinguish concatenation from layering. Resolve required merges with check_job.

Sound-effect contract

Use model: "sfx/elevenlabs-sound-v2".

• prompt describes the audible event or environment.
• duration_seconds is optional and accepts 0.5–30. Omit it when the natural event length is more important than exact timing.
• loop: true requests a seamless ambience or texture loop. Avoid one-shot attacks and resolved endings in a loop prompt.
• prompt_influence accepts 0–1. Higher values follow the wording more literally; lower values allow more variation. Start with the model default unless the result drifts from a specific cue.
• Use a supported output_format returned by get_model_params; the normal default is MP3.

Music contract

Use model: "music/elevenlabs-music-v1".

• prompt describes the musical result and is limited to 4,100 characters.
• music_length_ms accepts 3,000–600,000. Creative Claw defaults to 30,000 ms when omitted. Set it explicitly when the music must fit a video or edit.
• force_instrumental defaults to true in Creative Claw. Set it to false only when vocals are wanted, and describe the vocal role in the prompt.
• Prompt-only music can specify structure in prose. Do not invent a composition_plan field; it is not exposed by the current MCP tool.

Completion standard

Return the permanent audio asset and identify what was generated, the model, duration setting, and whether it loops or contains vocals. When pairing with video, distinguish the generated audio asset from the separately merged video result.