AI 编曲工作站 · AI Music Production Studio
纯浏览器专业编曲工具:真实乐器采样库 + Logic 风格 DAW + 多智能体 AI 作曲
A pro-grade, browser-only music workstation: real instrument samples + Logic-style DAW + multi-agent AI composition
💡 更新版本后若界面或按钮异常,请按 Ctrl / ⌘ + Shift + R 强制刷新一次。 浏览器可能缓存了旧版脚本;资源已加版本号,此后无需再手动强刷。
| 方式 | 操作 | 说明 |
|---|---|---|
| 零安装 | 直接双击 index.html |
采样从 CDN 下载,需联网 |
| 本地服务器(推荐) | 见下方命令 | 自带禁缓存头,更新永远即时生效 |
cd daw && python3 serve.py
# 浏览器打开 http://localhost:8000- 极速档(FluidR3 GM):22 种核心乐器 + 完整 GM 鼓组,每个半音一个真实采样,共 1266 个采样约 26 MB。开屏下载、Cache API 本地缓存,秒级进入。
- 高保真档(MusyngKite):进入后自动在后台把工程用到的乐器逐一升级为高保真采样(单文件体积约为极速档 2~4 倍),完成即无缝切换并释放旧档内存。顶栏「采样」徽标实时显示升级进度,可一键切换两档。
- 在线采样库:三套 GM 音源全集 + 打击乐镜像,总规模逾 10 GB,按需流式加载;另有 30+ 扩展乐器,选用时自动下载。
- 多轨编排视图 · 段落彩色时间轴
- 钢琴卷帘编辑器:鼓写谱模式、框选多选、复制粘贴、方向键移调、一键量化、人性化
- 混音台:每轨竖推子 + 实时电平表 + 三段 EQ + 混响发送 + 声像;Master 电平表与全局 Swing
- 卷积混响 + 总线压缩 + 母带砖墙限制器(防削波)
- 节拍器 · 循环播放 · 撤销/重做 · 自动保存
- 模型连接 — 只填 Base URL + 模型名 + API Key,即连任意平台(OpenAI / DeepSeek / Kimi / 智谱 / 通义 / OpenRouter / Groq / Ollama 本地 / 各类中转…)。协议自动识别(URL 含
anthropic走 Anthropic 协议,其余走 OpenAI 兼容),/v1路径缺漏自动补全;兼容思考型模型(DeepSeek-R1 / GLM / Qwen 思考模式等,思考耗尽输出上限时自动扩容重试)。密钥只存本机。 - 歌曲结构 — 像 Scratch 一样拖拽积木(前奏 / 主歌 / 预副歌 / 副歌 / 间奏 / 桥段 / 独奏 / 尾奏),也可一键套用风格预设或让 AI 推荐。
- 多智能体创作 — 1 位总制作人 Agent 定全曲蓝图(调性 / 速度 / 编制 / 核心动机 / 逐段和弦与能量),每个段落一个独立作曲家 Agent 顺序创作,携带上一段「结尾实录」保证前后衔接;System Prompt 明确约束连贯性、专业性、优美性、常规性四大铁律。
| 格式 | 用途 |
|---|---|
| WAV | 无损离线渲染 |
| MP3(320 kbps) | 发行分发 |
| 分轨 Stems | 每轨单独 WAV,供专业混音 / 发行交付 |
| 标准 MIDI | 跨 DAW 交换 |
工程文件(.mnvl.json) |
可再次打开继续编辑 |
| 按键 | 功能 |
|---|---|
空格 |
播放 / 暂停 |
Home |
回到开头 |
Ctrl/⌘ + Z |
撤销 |
Ctrl/⌘ + Shift + Z |
重做 |
钢琴卷帘:单击添加 · 拖拽移动 · 拖右缘改时值 · 右键 / Delete 删除 · Alt + 拖动改力度 · 滚轮上下 · Shift + 滚轮左右 · Ctrl + 滚轮缩放
- gleitz/midi-js-soundfonts(FluidR3 GM / MusyngKite,MIT)
- paulrosen/midi-js-soundfonts 镜像的打击乐
- jsDelivr 作备用镜像;MP3 编码采用 lamejs(LGPL,导出时按需加载)
💡 If the UI or buttons look broken after an update, press Ctrl / ⌘ + Shift + R once to force-refresh. Your browser may have cached old scripts; assets are now versioned, so no manual refresh will be needed afterwards.
| Method | How | Notes |
|---|---|---|
| Zero install | Double-click index.html |
Samples stream from CDN — internet required |
| Local server (recommended) | See command below | Ships with no-cache headers, so updates always take effect instantly |
cd daw && python3 serve.py
# Open http://localhost:8000 in your browser- Fast tier (FluidR3 GM) — 22 core instruments + full GM drum kit, one real sample per semitone: 1,266 samples, ~26 MB total. Downloads on launch, cached locally via the Cache API, ready in seconds.
- Hi-Fi tier (MusyngKite) — after launch, instruments used in your project are upgraded in the background to hi-fi samples (each file ~2–4× the fast tier's size), then swapped in seamlessly while old memory is released. A "Samples" badge in the top bar shows live upgrade progress; switch tiers with one click.
- Online sample library — three complete GM soundfont sets plus percussion mirrors, over 10 GB in total, streamed on demand; 30+ extra instruments auto-download when selected.
- Multi-track arrangement view · color-coded section timeline
- Piano-roll editor: drum notation mode, box multi-select, copy/paste, arrow-key transposition, one-click quantize, humanize
- Mixer: per-track vertical fader + live level meter + 3-band EQ + reverb send + pan; master meter and global swing
- Convolution reverb + bus compressor + mastering brickwall limiter (anti-clipping)
- Metronome · loop playback · undo/redo · autosave
- Model connection — just enter a Base URL + model name + API key to connect to any platform (OpenAI / DeepSeek / Kimi / Zhipu / Tongyi / OpenRouter / Groq / local Ollama / proxies…). The protocol is auto-detected (URLs containing
anthropicuse the Anthropic protocol, everything else uses OpenAI-compatible), missing/v1paths are auto-completed, and reasoning models are supported (DeepSeek-R1 / GLM / Qwen thinking modes, with automatic output-limit expansion and retry). API keys never leave your machine. - Song structure — drag blocks like Scratch (intro / verse / pre-chorus / chorus / interlude / bridge / solo / outro), apply a style preset in one click, or let the AI suggest one.
- Multi-agent composition — one Producer agent drafts the full-song blueprint (key / tempo / instrumentation / core motif / per-section chords and energy), then a dedicated Composer agent writes each section in sequence, carrying the previous section's actual ending to guarantee smooth transitions. The system prompt enforces four iron rules: coherence, professionalism, beauty, and convention.
| Format | Purpose |
|---|---|
| WAV | Lossless offline render |
| MP3 (320 kbps) | Release & distribution |
| Stems | One WAV per track, for pro mixing / delivery |
| Standard MIDI | Cross-DAW exchange |
Project file (.mnvl.json) |
Re-open and keep editing |
| Key | Action |
|---|---|
Space |
Play / pause |
Home |
Back to start |
Ctrl/⌘ + Z |
Undo |
Ctrl/⌘ + Shift + Z |
Redo |
Piano roll: click to add · drag to move · drag right edge to resize · right-click / Delete to remove · Alt + drag to change velocity · scroll vertically · Shift + scroll horizontally · Ctrl + scroll to zoom
- gleitz/midi-js-soundfonts (FluidR3 GM / MusyngKite, MIT)
- Percussion mirrored from paulrosen/midi-js-soundfonts
- jsDelivr as fallback mirror; MP3 encoding by lamejs (LGPL, loaded on demand at export time)
Made with 🎵 · MNVL Studio