Music Generation

Music Generation

Step 0 — plan first. Before generating, run the visual-planning skill (the pre-generation gate for all media, audio included): clarify the ask (duration, loop, where it plays), pin the intended mood/genre/tempo, decide what's IN/OUT (instrumentation to use or avoid), then draft and review the prompt.

Use the mcp__media-mcp__generate_music tool to create instrumental music via Google Gemini Lyria RealTime.

When to Use

• User asks to "create music", "generate a soundtrack", "make background music"
• User needs a jingle, intro music, ambient track, or sound design
• User wants music for a video, presentation, podcast, or app
• User describes a mood, genre, or musical style they want

Tool Reference

generate_music

Key parameters:

Weighted Prompts

The prompt system uses weights to blend musical concepts. Higher weight = more influence.

Single style

[{"text": "calm ambient piano with soft reverb", "weight": 1.0}]

Blended styles

[
  {"text": "electronic synthwave", "weight": 0.7},
  {"text": "orchestral strings", "weight": 0.3}
]

Genre + mood + instrumentation

[
  {"text": "jazz", "weight": 0.5},
  {"text": "relaxed late-night mood", "weight": 0.3},
  {"text": "saxophone and upright bass", "weight": 0.4}
]

Prompt Writing Guidelines

Describe three dimensions

1. Genre/style: jazz, electronic, orchestral, lo-fi, rock, ambient, cinematic
2. Mood/energy: calm, energetic, dark, uplifting, melancholic, suspenseful, triumphant
3. Instrumentation: piano, guitar, synth, strings, drums, bass, brass, woodwinds

Example prompts by use case

Background music for a product video:

[
  {"text": "modern corporate background music", "weight": 0.6},
  {"text": "upbeat and professional", "weight": 0.3},
  {"text": "light percussion and clean electric guitar", "weight": 0.3}
]

Podcast intro jingle:

[
  {"text": "short catchy podcast intro jingle", "weight": 0.7},
  {"text": "friendly and energetic", "weight": 0.4},
  {"text": "acoustic guitar and claps", "weight": 0.3}
]

Ambient background for coding:

[
  {"text": "lo-fi ambient chill beats", "weight": 0.6},
  {"text": "soft rain and vinyl crackle texture", "weight": 0.3},
  {"text": "slow tempo relaxing", "weight": 0.3}
]

Game soundtrack — boss battle:

[
  {"text": "epic orchestral battle music", "weight": 0.7},
  {"text": "intense fast-paced drums and brass", "weight": 0.5},
  {"text": "dark and dramatic", "weight": 0.3}
]

Meditation/relaxation:

[
  {"text": "peaceful meditation music", "weight": 0.6},
  {"text": "singing bowls and soft pads", "weight": 0.4},
  {"text": "very slow, spacious, minimal", "weight": 0.3}
]

BPM Guide

Scale Guide

Temperature Guide

Output Handling — MEDIA_OUTPUT_DIR

When MEDIA_OUTPUT_DIR is set (recommended): The MCP server saves the generated audio to a file and returns only the file path. Always use just this path — do NOT request or embed the raw audio data. This is critical because all MCP request/response messages are stored in the conversation history, and large base64 payloads pollute the context window, degrading performance.

When MEDIA_OUTPUT_DIR is not set: The MCP server has no choice but to return the audio as base64 data in the response. This works but is suboptimal for conversation history size.

Tips

• Start with a moderate temperature (0.5) and adjust based on results
• Use weighted prompts to blend genres rather than trying to describe everything in one text
• Specify BPM explicitly for rhythmic music — omit it for ambient/free-form
• Generated music is instrumental only — for vocals, combine with the speech generation skill
• For loopable background music, mention "seamless loop" in the prompt
• Generate multiple variations with different temperatures to find the best fit