Orca：67K stars 的并行 Agent 编排台，把 40+ CLI Agent 变成一支舰队

📌 开源仓库：stablyai/orca GitHub：https://github.com/stablyai/orca 官网：https://onOrca.dev License：MIT · Stars：67K+ 作者：Stably AI（YC W22） 创建：2026 年 3 月

---

BLUF：Orca 是一个让你同时调度多个 AI 编程 Agent 的桌面工具——每个 Agent 在独立的 git worktree 里跑，互不干扰，结果并排对比。它不替你订阅 Claude Code 或 Codex，而是用你自己的订阅，把这些 Agent 变成一支可以统一指挥的舰队。MIT 开源，67K stars，支持 40+ 主流 CLI Agent，macOS / Windows / Linux 全平台 + iOS/Android 手机端。

---

问题是什么

每个 AI 编程 Agent 都有自己的操作方式。Claude Code 有 Claude Code 的 keybinding，Codex 有 Codex 的上下文管理，Cline 有 Cline 的工具调用接口。

当你同时用三四个 Agent 处理不同任务时，你需要：

• 来回切终端窗口
• 手动管理哪个 Agent 在哪个分支上
• 盯着四个不同的输出流想”这个搞完了没有”
• 从床上爬起来看看跑了两小时的那个 Agent 到底成没成

Orca 解决的就是这些摩擦。

---

核心设计：Parallel Worktrees

Orca 最核心的能力是 Parallel Worktrees：

操作流程：

1. 给出一个任务（比如”给这个 API 加速率限制”）
2. Orca 把这个任务扇出给多个 Agent（比如 Claude Code × 2、Codex × 1）
3. 每个 Agent 在独立的 git worktree 里跑，互不影响主分支和彼此
4. 任务跑完后，并排对比三份 diff
5. 选最好的那份合并进主分支

这不只是”多开几个终端”。worktree 隔离意味着：Agent 可以随意修改文件、运行测试、甚至制造错误，不会污染你的工作区。对比完毕后，不满意的直接丢掉——零恢复成本。

---

终端层：Ghostty 级别的渲染

Orca 的终端不是普通的嵌入 xterm。

技术栈：WebGL 渲染，和 Ghostty 同档次。支持：

• 无限分屏（横/竖任意组合）
• Scrollback 跨应用重启持久化
• 所有 Agent 的输出都在同一个界面里，不需要 Cmd+Tab 切换

对于”盯着 Agent 跑”这件事，终端体验直接影响注意力成本——这也是 Orca 花精力做终端而不只做 orchestration 的理由。

---

Design Mode：点 UI 元素，直喂 Agent

这是 Orca 里最有意思的功能之一。

原理：Orca 内嵌了真实的 Chromium 浏览器（不是 webview）。在 Design Mode 下，你点击页面上任意一个 UI 元素，Orca 会自动提取：

• 该元素的 HTML 结构
• 相关 CSS 样式
• 一张裁剪好的截图

这些内容会作为上下文直接注入到你的 Agent prompt 里。

实际用途：你不需要再手写”那个蓝色按钮的 padding 太大了”——你直接点那个按钮，Agent 拿到的是精确的 DOM 结构和样式，而不是你描述的模糊语言。

---

GitHub & Linear 原生集成

Orca 不只是终端管理器，它把任务管理也集成了进来：

• GitHub：应用内浏览 PR、issues、项目看板，从任意 issue 直接开一个 worktree
• Linear：同样的逻辑，Linear ticket → worktree，无需切应用

这意味着从”看到一个 bug” 到”Agent 开始修它”的流程，全部在 Orca 里完成，不需要浏览器、不需要手动 git checkout -b。

---

SSH Worktrees：在远端高性能机器跑 Agent

本地 MacBook 跑 Agent 没问题，但如果要同时跑五个，或者跑需要大量计算的任务呢？

Orca 的 SSH Worktrees：

• 在远端 Linux 服务器（甚至 GPU 机器）上建立 worktree
• 完整的文件编辑、git 操作、终端访问
• 自动重连：网络断了，Orca 自动恢复 session
• 端口转发：远端跑的 dev server 直接在本地访问

这让”Agent 跑在算力更大的地方”成为日常操作，而不是需要配置 tmux + rsync 的麻烦事。

---

Mobile Companion：手机上调度 Agent

Orca 有 iOS 和 Android 配套 App（iOS 在 App Store，Android 提供 APK）。

手机端的能力：

• 实时监控所有 Agent 的运行状态
• 收到 Agent 完成的推送通知
• 发送后续指令（follow-up prompt）

这直接解决了”晚上让 Agent 跑，不知道什么时候完成”的问题。Agent 跑完了，手机推送一条通知，睡觉前扫一眼，满意就 merge，不满意明天再说。

---

支持的 Agent 列表（40+）

Orca 不绑定任何 Agent，列表包括：

主流编程 Agent：Claude Code、Codex（OpenAI）、OpenCode、Pi、Cline、Charm（charmbracelet/crush）

新兴选手：Codebuff、Command Code、Continue、Droid（Factory AI）、Kilocode、Kimi Code、Kiro（AWS）、Mistral Vibe、Qwen Code、Rovo Dev（Atlassian）

以及”any CLI agent”——只要是命令行接口，Orca 理论上都能跑。

---

安装

桌面端（macOS / Windows / Linux）：

# macOS (Homebrew)
brew install --cask stablyai/orca/orca

# Arch Linux
yay -S stably-orca-bin

# 直接下载：https://onorca.dev/download

手机端：

• iOS：App Store 搜 “Orca IDE” 或 TestFlight
• Android：GitHub releases 下载 APK

---

几个值得关注的设计决策

BYOK（Bring Your Own Key）：Orca 不收订阅费，你用自己的 Claude/OpenAI 等订阅跑 Agent。这是重要的定位——它是工具层，不是中间商。

MIT 开源：代码完全开放，可以自己 build，也可以二次开发。云端配对服务（mobile companion 用到的 relay）的代码也在仓库的 cloud/ 目录下。

Stably AI 是谁：YC W22 背景，最早做 stablecoin 基础设施，Orca 是他们 2026 年的新方向。67K stars 是 2026 年 3 月到 9 月的增长——六个月内从零到这个量级，说明这个方向的需求真实存在。

---

开源代码与模型仅供学习、勿直接用于工作。

© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用，须注明作者姓名及原文链接，不得去除署名后以原创发布。

📌 Repository: stablyai/orca GitHub: https://github.com/stablyai/orca Website: https://onOrca.dev License: MIT · Stars: 67K+ Author: Stably AI (YC W22) Created: March 2026

---

BLUF: Orca is a desktop tool for orchestrating multiple AI coding agents simultaneously — each agent runs in an isolated git worktree, results sit side-by-side for comparison. It doesn’t replace your Claude Code or Codex subscription; it uses your existing subscriptions and turns those agents into a coordinated fleet. MIT-licensed, 67K stars, supports 40+ CLI agents, macOS/Windows/Linux plus iOS/Android mobile companion.

---

The Problem

Every AI coding agent has its own interface. Claude Code has its own keybindings, Codex its own context management, Cline its own tool-call surface.

Running three or four agents on different tasks simultaneously means:

• Constant terminal window switching
• Manually tracking which agent is on which branch
• Watching four output streams wondering “is that one done yet?”
• Getting out of bed to check whether the two-hour run finished

Orca eliminates that friction.

---

Core Design: Parallel Worktrees

Orca’s most central capability is Parallel Worktrees:

Flow:

1. Describe a task (e.g. “add rate limiting to this API”)
2. Orca fans that task out to multiple agents (e.g. Claude Code × 2, Codex × 1)
3. Each agent runs in an isolated git worktree — no interference with your working branch or each other
4. When tasks finish, compare three diffs side by side
5. Pick the best one and merge

This isn’t “open a few terminals.” Worktree isolation means agents can freely modify files, run tests, even make mistakes — without contaminating your workspace. Discard what you don’t want, zero recovery cost.

---

Terminal Layer: Ghostty-Class Rendering

Orca’s terminal isn’t a basic embedded xterm.

Stack: WebGL rendering, same tier as Ghostty. Features:

• Unlimited splits, any horizontal/vertical arrangement
• Scrollback persists across app restarts
• All agent output in one interface — no Cmd+Tab required

Terminal quality directly affects attention cost when watching agents work. That’s why Orca invested in the terminal layer rather than just the orchestration.

---

Design Mode: Click UI Elements, Feed Them to Your Agent

One of Orca’s most interesting features.

How it works: Orca embeds a real Chromium browser (not a webview). In Design Mode, click any UI element and Orca automatically extracts: the HTML structure, the relevant CSS, and a cropped screenshot.

This context is injected directly into your agent prompt.

Practical use: Instead of writing “that blue button has too much padding,” you click the button. The agent receives the exact DOM structure and styles — not your approximation of them.

---

Native GitHub & Linear Integration

Orca integrates task management directly:

• GitHub: Browse PRs, issues, and project boards in-app — open a worktree directly from any issue
• Linear: Same logic — Linear ticket → worktree, no app switching needed

From “spotted a bug” to “agent is working on it” — the entire flow stays inside Orca.

---

SSH Worktrees: Run Agents on Remote Hardware

Local laptop is fine for one agent. Five concurrent agents, or compute-heavy tasks?

Orca’s SSH Worktrees let you:

• Establish worktrees on a remote Linux server (including GPU machines)
• Full file editing, git operations, terminal access
• Auto-reconnect: network drops, Orca recovers the session automatically
• Port forwarding: dev servers running remotely are directly accessible locally

Running agents on beefier hardware becomes a routine operation rather than a tmux + rsync configuration exercise.

---

Mobile Companion: Steer Agents from Your Phone

Orca has iOS and Android companion apps (iOS on App Store, Android APK).

Mobile capabilities:

• Real-time monitoring of all running agents
• Push notifications when agents complete
• Send follow-up prompts from anywhere

This solves “let agents run overnight, don’t know when they finish.” Agent completes → push notification → check from bed → merge if satisfied, revisit in the morning if not.

---

Supported Agents (40+)

Orca isn’t bound to any single agent. The list includes:

Established coding agents: Claude Code, Codex (OpenAI), OpenCode, Pi, Cline, Charm

Newer entrants: Codebuff, Command Code, Continue, Droid (Factory AI), Kilocode, Kimi Code, Kiro (AWS), Mistral Vibe, Qwen Code, Rovo Dev (Atlassian)

Plus “any CLI agent” — if it has a command-line interface, Orca can run it.

---

Install

Desktop (macOS / Windows / Linux):

# macOS (Homebrew)
brew install --cask stablyai/orca/orca

# Arch Linux
yay -S stably-orca-bin

# Direct download: https://onorca.dev/download

Mobile:

• iOS: App Store (“Orca IDE”) or TestFlight
• Android: APK from GitHub releases

---

Design Decisions Worth Noting

BYOK (Bring Your Own Key): Orca charges no subscription fee — you run agents with your own Claude/OpenAI subscriptions. It’s tooling, not a middleman.

MIT-licensed: Fully open-source, buildable from source, forkable. The mobile companion relay service code is also in the repo under cloud/.

Who is Stably AI: YC W22 alumni, originally stablecoin infrastructure, Orca is their 2026 direction. 67K stars in six months (March to September 2026) signals that the underlying demand is real.

---

Open-source code and models are for learning purposes only — do not use directly in production work.

© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.