midi-composer-mcp
Provides atomic music-theory and MIDI tools for composing, enabling LLMs to chain deterministic steps like scale/chord lookups, degree resolution, rhythm generation, and MIDI rendering.
README
midi-composer-mcp
An MCP server that gives an LLM atomic music-theory and MIDI tools for composing.
The design principle: the tools contain no creativity. Every tool is a small, deterministic (or seeded-random) step — looking up scales and chords, matching notes, resolving degree sequences, rolling dice, rendering MIDI. The LLM is the composer: it chains the tools, makes every musical choice, and the tools do the mechanical work correctly.
All tools are compatible with each other: notes, chord symbols and rhythm patterns returned by one tool are valid inputs to every other tool.
Note format
- Notes are strings:
C,F#,Bb,Ebb(case-insensitive, unicode♯/♭accepted). - A note without an octave is an abstract pitch class.
- A note with an octave is a concrete pitch:
C4is middle C (MIDI 60),Eb3,A5... Generation respects it:get_scale("major", "C5")→C5 D5 E5 F5 G5 A5 B5 C6with MIDI numbers;get_chord("9", "C4")→C4 E4 G4 Bb4 D5. - Note lists may be JSON arrays (
["C", "E", "G"]) or plain strings ("c e g","C, E, G"). - Spelling is proper: F major has a
Bb(notA#), Cdim7 has aBbb.
Tools
| Tool | What it does (deterministic unless noted) |
|---|---|
list_scales |
Scale-type database: intervals, degree labels, aliases. |
list_chords |
Chord-type database: intervals, degrees, symbol suffixes, aliases. |
get_scale |
Describe a scale type; with a root, generate its notes. maj + C → C D E F G A B C. |
get_chord |
Describe a chord type; with a root, generate its notes. min + F → F Ab C. |
match_scales |
Scales containing the given notes (octave-insensitive), exact matches flagged. |
match_chords |
Chords matching the given notes; c e g → C, e g c → C/E (first inversion), partial matches list missing notes. |
diatonic_chords |
The chord on each scale degree, with roman numerals, degree names and harmonic functions (tonic/subdominant/dominant). |
degrees_to_chords |
Resolve a degree sequence you chose ([1,5,6,4], "I V vi IV", "1-5-6-4") into concrete chords. |
random_notes |
🎲 Uniform random picks from any note pool (seeded, reproducible). |
random_rhythm |
🎲 Random rhythm pattern like O...Oo.. — O strong beat, o weak beat, . pause (seeded, reproducible). |
notes_to_midi |
Render a note sequence (scale, arpeggio, melody) to a .mid file, optionally shaped by a rhythm pattern. |
chords_to_midi |
Render chord symbols and/or note arrays to a .mid file (block or arpeggiated). |
song_to_midi |
Render melody + chord accompaniment into one two-track .mid file. |
MIDI tools write to ./midi_output (override per call with output_dir or globally with the MIDI_COMPOSER_OUTPUT_DIR environment variable) and also return the file base64-encoded, plus the exact events written.
A composing session looks like this
The LLM drives; each tool call is one mechanical step:
get_scale("minor", "A")→A B C D E F G Adiatonic_chords("A", "minor")→i=Am, ii°=Bdim, bIII=C, iv=Dm, v=Em, bVI=F, bVII=G(+ harmonic functions)- LLM decides on
i–VI–III–VII→degrees_to_chords("A", "minor", "i VI III VII")→Am F C G random_rhythm(length=32, seed=42)→.Oo..oOOoOoo.O.Oo...oo...oOOOooOrandom_notes(notes=<scale notes>, count=21, seed=42)→ melody material (or the LLM writes the melody itself)song_to_midi(melody_notes=..., melody_rhythm=..., chords=["Am","F","C","G"])→song.mid
Every intermediate result is plain data the LLM can inspect, edit by hand (e.g. tweak the rhythm string), or feed into another tool — e.g. get_chord("m7","A") → match_scales(...) to find scales for a solo.
Installation
Requires Python ≥ 3.10.
# with uv (recommended)
uv pip install . # or: uv sync && uv run midi-composer-mcp
# or with pip
pip install .
Run the server (stdio transport):
midi-composer-mcp
# or without installing:
uv run --with mcp --with mido python -m midi_composer_mcp.server
Claude Code
claude mcp add midi-composer -- uv run --directory /path/to/midi-composer-mcp midi-composer-mcp
Claude Desktop
{
"mcpServers": {
"midi-composer": {
"command": "uv",
"args": ["run", "--directory", "/path/to/midi-composer-mcp", "midi-composer-mcp"],
"env": { "MIDI_COMPOSER_OUTPUT_DIR": "/path/to/your/midi/files" }
}
}
}
Development
uv venv && uv pip install -e ".[dev]"
.venv/bin/python -m pytest
Layout:
src/midi_composer_mcp/
notes.py # note parsing, proper spelling, octaves, MIDI numbers
scales.py # scale database, generation, matching
chords.py # chord database, symbols, generation, matching
diatonic.py # chords per scale degree, degree-sequence resolution
generate.py # seeded dice: random notes, random rhythm
midi_io.py # deterministic MIDI rendering (mido)
server.py # the MCP server (FastMCP) — thin wrappers over the above
Roadmap ideas
- Transposition and inversion helpers
- Humanize options (timing/velocity jitter as a seeded, mechanical step)
- Reading MIDI files back into note/chord data
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。