O
oarthurbg/orchestra
Many voices. One composition. A Claude Code plugin for running one project with an orchestrator session and…
在 GitHub 查看 ↗链接已于 2026-11-12 核查 · HTTP 200 · 打开的是厂商自己的官网
安装与配置
客户端配置（取自该仓库的安装命令）
{
"mcpServers": {
"oarthurbg-orchestra": {
"command": "npx",
"args": ["orchestra-install"]
}
}
}解析自我们于 2026-11-03 抓取的该仓库 README；我们未执行过。
command/args 来自仓库写明的安装命令；不确定时请以该项目 README 为准。
VS Code（.vscode/mcp.json —— 键名是 servers）
{
"servers": {
"oarthurbg-orchestra": {
"command": "npx",
"args": ["orchestra-install"]
}
}
}VS Code 顶层键是 <code>servers</code>，其余与上面那份相同。
Claude Code（命令行）
claude mcp add --transport stdio oarthurbg-orchestra -- npx orchestra-install
Claude Code 不收 JSON 文件，用这条命令直接加。
配置项的键名以你的客户端为准：客户端与 CLI
- ★ 数取值
- 2 ★ · oarthurbg/orchestra · 取自 GitHub 搜索 API · （2026-11-13 20:00:50 核对）
- 来源明细
- 从该仓库的 GitHub 页面直接读取 · 条目类型：开源项目
- 收录与活跃
- 收录于 2026-11-04 · 最近见到 2026-11-14 · 共见到 11 次 · 被 1 个来源列出
- 评分构成
- 综合分 103 · 2 ★ · 1 个来源 · repo
- 安装
npx orchestra-install
常见问题（答案都来自我们的记录）
怎么安装？
该仓库给出的安装命令是
npx orchestra-install。我们的记录里它是仓库自己提供的；不保证在你的客户端上直接可用。它适合谁 / 用来做什么？
我们把它归在这些标签下: 开源项目。标签来自条目自身，不是我们强加的分类。
谁维护它、还在更新吗？
仓库
oarthurbg/orchestra；我们收录于 2026-11-04；最近见到 2026-11-14。这里没有"官方认证"或"质量评级"——我们只记录来源与时间。# orchestra
以下摘录自该仓库的 README（片段） · 在 GitHub 读完整 README →
Run one project with many Claude Code sessions at once.
One session plans. One session per feature builds. You make the decisions, and every change is reviewed before it merges.
orchestra is a Claude Code plugin, built while developing an internal media-planning app. It works for any kind of app, new or already under way: no database, server or Docker required.
## How it works
1. You and the maestro plan the app. It splits the work into features and opens one session for each, as many at once as your machine and your plan limits allow.
2. Each principal builds one feature in its own worktree: it mocks up and specs with you, then sends subagents to plan, build, review and test.
3. Every feature ends as a pull request. The maestro checks it and merges it. When a session needs a decision, it asks you.
## Who does what
Wherever orchestra draws a session, its shape tells you the role and its fill tells you the state: solid is running, outlined is idle, over a heavy rule is waiting on you. Agents has the nine subagents.
## See it
The band sits above the prompt in every session: your plan limits, when the week will run out, and who needs you, most urgent first. It also offers new versions of orchestra. More about the band.
The liveview is a page with every workstream's progress, ETA and timeline. It runs on your machine, uses no tokens, and opens on your phone too, over Tailscale. More about the liveview.
The inbox lists every question waiting on you, oldest first, with the mockups and specs you need to answer it.
## Why orchestra
Ship features in parallel. One session and one worktree per feature, as many as your machine and plan allow. Read more
The right model for each job. Nine subagents: Opus for judgment, cheaper models for chores. Read more
A proven process, built in. Brainstorm, spec, plan, build, review, PR, with the best skill at each stage. Read more
Nothing merges unreviewed. Spec, quality and finish reviews, then the whole branch. No session merges its own PR. Read more
No surprises with your limits. The band shows what is left, when the week runs out, and who needs you. Read more
See everything from anywhere. Progress, ETAs and your questions, on your phone too. No tokens. Read more
Fewer tokens to find code. A language server instead of grep: up to 65% less, measured. Read more
Long sessions stay sharp. A handoff note and a clear past 250k tokens; nothing lost on a compact. Read more
Cheaper waits. The prompt cache stays warm while subagents run. Read more
Safe to leave running. You set what sessions may do alone; a branch guard blocks direct pushes. Read more
## Does it help?
Measured, not estimated: every number comes from a script in bench/, explained in Benchmarks.
Code navigation saves tokens. Finding who calls what with a language server instead of grep reaches the same answers with fewer tokens, most of all for small models on big codebases.
One session plans. One session per feature builds. You make the decisions, and every change is reviewed before it merges.
orchestra is a Claude Code plugin, built while developing an internal media-planning app. It works for any kind of app, new or already under way: no database, server or Docker required.
## How it works
1. You and the maestro plan the app. It splits the work into features and opens one session for each, as many at once as your machine and your plan limits allow.
2. Each principal builds one feature in its own worktree: it mocks up and specs with you, then sends subagents to plan, build, review and test.
3. Every feature ends as a pull request. The maestro checks it and merges it. When a session needs a decision, it asks you.
## Who does what
Wherever orchestra draws a session, its shape tells you the role and its fill tells you the state: solid is running, outlined is idle, over a heavy rule is waiting on you. Agents has the nine subagents.
## See it
The band sits above the prompt in every session: your plan limits, when the week will run out, and who needs you, most urgent first. It also offers new versions of orchestra. More about the band.
The liveview is a page with every workstream's progress, ETA and timeline. It runs on your machine, uses no tokens, and opens on your phone too, over Tailscale. More about the liveview.
The inbox lists every question waiting on you, oldest first, with the mockups and specs you need to answer it.
## Why orchestra
Ship features in parallel. One session and one worktree per feature, as many as your machine and plan allow. Read more
The right model for each job. Nine subagents: Opus for judgment, cheaper models for chores. Read more
A proven process, built in. Brainstorm, spec, plan, build, review, PR, with the best skill at each stage. Read more
Nothing merges unreviewed. Spec, quality and finish reviews, then the whole branch. No session merges its own PR. Read more
No surprises with your limits. The band shows what is left, when the week runs out, and who needs you. Read more
See everything from anywhere. Progress, ETAs and your questions, on your phone too. No tokens. Read more
Fewer tokens to find code. A language server instead of grep: up to 65% less, measured. Read more
Long sessions stay sharp. A handoff note and a clear past 250k tokens; nothing lost on a compact. Read more
Cheaper waits. The prompt cache stays warm while subagents run. Read more
Safe to leave running. You set what sessions may do alone; a branch guard blocks direct pushes. Read more
## Does it help?
Measured, not estimated: every number comes from a script in bench/, explained in Benchmarks.
Code navigation saves tokens. Finding who calls what with a language server instead of grep reaches the same answers with fewer tokens, most of all for small models on big codebases.
收录徽章（可选）
想把收录标记放进 README 就复制这行，可选，不影响收录。
[](https://anclum.com/zh/s/oarthurbg-orchestra/)
- 来源
- github · https://github.com/oarthurbg/orchestra
- 按领域浏览
- 开源项目
- Traction
- 2 ★
- 收录于
- 今天
同类推荐
所属领域与标签
每个标签都能点进去看同类的全部条目。数字是当前库里的真实条数。