"The best way to predict the future is to invent it." — Alan Kay
「預測未來最好的方式就是發明它。」—— Alan Kay
Why CLI Agents, Not API Calls?
There's a fundamental architectural split in the AI coding agent world:
- API-based orchestration: Your code calls the Claude/GPT API directly, manages context windows, handles tool use, and implements agent loops. Frameworks like CrewAI, AutoGen, and LangGraph live here.
- CLI-based orchestration: You spawn actual CLI processes —
claude,codex,gemini— that run as autonomous agents with their own context, tools, and shell access. You manage processes, not prompts.
The CLI approach has a crucial advantage: the agent runtime handles everything — context management, tool use, file editing, git operations, test execution. The orchestrator only needs to answer: what task?, which agent?, when to start?, when to stop?
This guide surveys every notable open-source project in the CLI-based orchestration space as of March 2026. Over 30 projects, organized into 9 categories.
為什麼是 CLI Agent，而不是 API 呼叫？
AI coding agent 的世界有一個根本的架構分歧：
- API 型編排： 你的程式碼直接呼叫 Claude/GPT API，管理 context window、處理 tool use、實作 agent 迴圈。CrewAI、AutoGen、LangGraph 屬於這一類。
- CLI 型編排： 你啟動實際的 CLI 程序——
claude、codex、gemini——它們作為自主 agent 運行，有自己的 context、工具和 shell 存取權限。你管理的是程序，不是 prompt。
CLI 方式有一個關鍵優勢：agent runtime 處理一切——context 管理、tool use、檔案編輯、git 操作、測試執行。編排器只需要回答：什麼任務？、哪個 agent？、何時啟動？、何時停止？
本指南盤點截至 2026 年 3 月，CLI 型編排領域中每一個值得關注的開源專案。超過 30 個專案，橫跨 9 大類別。
The Landscape at a Glance
| Category | Key Projects | Stars Range |
|---|---|---|
| Business orchestration | Paperclip | 27.6k |
| Claude Code-specific | oh-my-claudecode, Gas Town, Multiclaude, agents (wshobson) | 500–31.5k |
| Multi-agent platforms | Ruflo, Agent Orchestrator, Overstory, MyClaude | 1k–21.5k |
| Session managers | Claude Squad, dmux, cmux, Mux | 1.2k–7.5k |
| Codex CLI-specific | Codex Orchestrator, Symphony | 235–13.1k |
| Desktop orchestrators | Aperant, Dorothy, Jean, AI Maestro | 211–13.3k |
| Sandboxed execution | TSK, Agent Swarm | 150–267 |
| Lightweight / skill-based | Swarms, Sudocode, Claude MPM | 85–232 |
| Mobile / remote | Happy Coder | 15.5k |
Dominant patterns across all projects:
- tmux + git worktrees — the standard isolation mechanism
- Mayor/Lead/Supervisor — a coordinator agent decomposes tasks and delegates
- Wave-based execution — independent tasks run in parallel, dependent tasks wait
- SQLite or git-backed state — lightweight persistence without external databases
全景一覽
| 類別 | 代表專案 | 星數範圍 |
|---|---|---|
| 商業編排 | Paperclip | 27.6k |
| Claude Code 專用 | oh-my-claudecode、Gas Town、Multiclaude、agents (wshobson) | 500–31.5k |
| 多 agent 平台 | Ruflo、Agent Orchestrator、Overstory、MyClaude | 1k–21.5k |
| Session 管理器 | Claude Squad、dmux、cmux、Mux | 1.2k–7.5k |
| Codex CLI 專用 | Codex Orchestrator、Symphony | 235–13.1k |
| 桌面編排器 | Aperant、Dorothy、Jean、AI Maestro | 211–13.3k |
| 沙箱執行 | TSK、Agent Swarm | 150–267 |
| 輕量 / skill 型 | Swarms、Sudocode、Claude MPM | 85–232 |
| 行動 / 遠端 | Happy Coder | 15.5k |
所有專案中的主流模式：
- tmux + git worktrees —— 標準的隔離機制
- Mayor/Lead/Supervisor —— 協調者 agent 分解任務並委派
- 波次執行（Wave-based execution） —— 獨立任務平行執行，有相依性的任務等待
- SQLite 或 git-backed 狀態 —— 輕量持久化，不需要外部資料庫
Category 1: Business Orchestration — Paperclip
Paperclip | 27.6k stars | MIT | Node.js + React + PostgreSQL
The "zero-human company" platform. Paperclip models an entire business: org charts, reporting lines, goal hierarchies, per-agent budgets with auto-pause, heartbeat scheduling, board-level governance with approval gates, and immutable audit logs. It supports 9+ agent adapters (Claude Code, Codex, OpenClaw, Cursor, Gemini, and more).
Unique to Paperclip:
- Multi-company support with complete data isolation
- Atomic task checkout (no two agents work the same issue)
- Budget enforcement with soft alerts (80%) and hard ceilings (100%)
- Board approval gates for strategy, hires, and budget changes
Best for: Founders building "AI companies" — multiple agents working toward a business goal with governance.
Limitation: Heavy for solo developers who just want parallel coding agents.
類別 1：商業編排——Paperclip
Paperclip | 27.6k 星 | MIT | Node.js + React + PostgreSQL
「零人公司」平台。Paperclip 建模整間公司：組織架構、報告線、目標層級、每 agent 預算（含自動暫停）、heartbeat 排程、董事會層級的治理（含核准閘門）、以及不可變的稽核日誌。支援 9 種以上 agent adapter（Claude Code、Codex、OpenClaw、Cursor、Gemini 等）。
Paperclip 獨有：
- 多公司支援，完全資料隔離
- 原子任務 checkout（沒有兩個 agent 同時處理同一個 issue）
- 預算控制：80% 軟警報 + 100% 硬上限自動暫停
- 董事會核准閘門（策略、僱用、預算變更）
最適合： 建構「AI 公司」的創辦人——多個 agent 朝向商業目標工作，並有治理機制。
限制： 對只想要平行 coding agent 的單人開發者來說太重。
Category 2: Claude Code-Specific Orchestration
wshobson/agents — 31.5k stars
agents | MIT | The largest project by stars
Not a standalone orchestrator but a massive plugin ecosystem for Claude Code: 112 specialized agents, 16 multi-agent workflow orchestrators, 146 skills, and 79 dev tools organized into 72 Claude Code plugins. Uses 4-tier model assignment (Opus for architecture, Sonnet for development, Haiku for operations). Installs via /plugin marketplace.
Best for: Developers who want to supercharge Claude Code without leaving its native environment.
Gas Town — 12.4k stars
Gas Town | Go | By Steve Yegge
A deeply git-native orchestration system. A "Mayor" coordinator manages 20-30 agents using git-backed persistent state: "beads" for issue tracking, "convoys" for work bundles, "hooks" for agent state. Supports 10+ runtimes: Claude, Gemini, Codex, Cursor, Auggie, Amp, OpenCode, Copilot, Pi, OMP. Features formula workflows (TOML-defined repeatable processes), TUI activity feeds, and problem detection for stuck agents.
Best for: Teams that want orchestration deeply integrated with git workflows.
Key difference from Paperclip: Git is the source of truth, not PostgreSQL.
oh-my-claudecode — 10.1k stars
oh-my-claudecode | npm: oh-my-claude-sisyphus
Teams-first multi-agent orchestration with 32 specialized agents and smart model routing (Haiku for simple tasks, Opus for complex). 9 execution modes including Ultrawork (maximum parallelism) and Ralph (persistent verify/fix loops). Supports Claude Code natively, plus Codex/Gemini via tmux workers. The /ccg command synthesizes analysis from all three providers. Claims 30-50% token cost reduction through intelligent routing.
Best for: Power users who want sophisticated model routing and cost optimization.
Multiclaude — 504 stars
Multiclaude | Go | MIT
Spawns autonomous Claude Code agents in tmux windows with git worktrees. Uses a "Brownian Ratchet" philosophy — controlled chaos with CI/testing as the filtering mechanism. Built-in agent types: Supervisor, Merge Queue, PR Shepherd, Worker, Reviewer. Single-player mode auto-merges on CI pass; multiplayer mode respects team review processes.
Best for: Developers who trust CI as the quality gate and want agents to "throw code at the wall and see what sticks."
SwarmSDK (parruda/claude-swarm) — 1.7k stars
SwarmSDK | Ruby
A Ruby framework for multi-agent orchestration with persistent memory (FAISS-indexed semantic search). Single-process model using RubyLLM for direct method calls. Node workflows with dependencies, 12-event hooks system.
Best for: Ruby developers who want a programming framework, not a deployment platform.
Claude Code Agent Teams — Official Feature
Built into Claude Code v2.1.32+. Enable with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS. One session acts as team lead, coordinating work and synthesizing results. Teammates work independently with their own context windows. This is the foundation many third-party tools build on or compete with.
類別 2：Claude Code 專用編排
wshobson/agents——31.5k 星
agents | MIT | 星數最高的專案
不是獨立的編排器，而是 Claude Code 的大型 plugin 生態系：112 個專門 agent、16 個多 agent 工作流編排器、146 個 skill、79 個開發工具，組織成 72 個 Claude Code plugin。使用 4 層模型分配（Opus 負責架構、Sonnet 負責開發、Haiku 負責維運）。透過 /plugin marketplace 安裝。
最適合： 想在 Claude Code 原生環境內強化功能的開發者。
Gas Town——12.4k 星
Gas Town | Go | Steve Yegge 作品
深度 git 原生的編排系統。「Mayor」協調者使用 git-backed 持久化狀態管理 20-30 個 agent：「beads」追蹤 issue、「convoys」打包工作、「hooks」記錄 agent 狀態。支援 10+ 種 runtime：Claude、Gemini、Codex、Cursor、Auggie、Amp、OpenCode、Copilot、Pi、OMP。具備 formula workflow（TOML 定義的可重複流程）、TUI activity feed、以及卡住 agent 的偵測。
最適合： 想要編排與 git 工作流深度整合的團隊。
與 Paperclip 的關鍵差異： Git 是真相來源，不是 PostgreSQL。
oh-my-claudecode——10.1k 星
oh-my-claudecode | npm: oh-my-claude-sisyphus
以團隊為核心的多 agent 編排，有 32 個專門 agent 和智慧模型路由（Haiku 處理簡單任務、Opus 處理複雜任務）。9 種執行模式，包括 Ultrawork（最大平行度）和 Ralph（持續驗證/修復迴圈）。原生支援 Claude Code，另透過 tmux worker 支援 Codex/Gemini。/ccg 指令能綜合三個 provider 的分析。宣稱透過智慧路由減少 30-50% token 成本。
最適合： 想要精密模型路由和成本優化的 power user。
Multiclaude——504 星
Multiclaude | Go | MIT
在 tmux 視窗中搭配 git worktree 啟動自主 Claude Code agent。採用**「布朗棘輪」哲學**——受控的混沌，以 CI/測試作為篩選機制。內建 agent 類型：Supervisor、Merge Queue、PR Shepherd、Worker、Reviewer。單人模式在 CI 通過時自動 merge；多人模式尊重團隊 review 流程。
最適合： 信任 CI 作為品質閘門、想讓 agent「把程式碼往牆上丟看什麼能黏住」的開發者。
SwarmSDK（parruda/claude-swarm）——1.7k 星
SwarmSDK | Ruby
Ruby 多 agent 編排框架，具備持久化記憶（FAISS 索引的語意搜尋）。單程序模型，使用 RubyLLM 直接方法呼叫。節點工作流有相依性、12 個事件 hook。
最適合： 想要程式框架而非部署平台的 Ruby 開發者。
Claude Code Agent Teams——官方功能
內建於 Claude Code v2.1.32+。透過 CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS 啟用。一個 session 作為 team lead，協調工作並綜合結果。Teammate 在自己的 context window 中獨立工作。這是許多第三方工具的基礎或競爭對象。
Category 3: Multi-Agent Platforms (Agent-Agnostic)
Ruflo — 21.5k stars
Ruflo | MIT | Formerly "Claude Flow"
The most feature-rich orchestration platform. 60+ specialized agents organized into swarms led by "queens" (Strategic, Tactical, Adaptive). Features: Q-Learning router, Mixture-of-Experts with 8 experts, 42+ skills, multiple consensus algorithms (Byzantine, Raft, Gossip, CRDT), WASM agent booster (<1ms for simple transforms), multi-provider support (Claude, GPT, Gemini, Cohere, Ollama). Claims ~250% extended Claude Code usage and 30-50% token optimization.
Best for: Teams wanting maximum feature breadth.
Caveat: The claims around self-learning and consensus algorithms warrant careful evaluation.
Agent Orchestrator (ComposioHQ) — 4.5k stars
Agent Orchestrator | MIT | TypeScript
The cleanest architecture in this space. 8 swappable plugin slots, each defined by a TypeScript interface:
| Slot | Options |
|---|---|
| Runtime | tmux / Docker / K8s |
| Agent | Claude Code / Codex / Aider / OpenCode |
| Workspace | worktree / clone |
| Tracker | GitHub / Linear |
| SCM | Git |
| Notifier | Slack / webhook |
| Terminal | tmux |
| Lifecycle | hooks |
Agents autonomously fix CI failures, address review comments, and open PRs. 3,288 test cases.
Best for: Teams that value clean architecture and extensibility.
Overstory — 1k stars
Overstory | MIT | TypeScript/Bun
The most runtime-agnostic orchestrator. Supports 8 agent runtimes: Claude Code, Sapling, Pi, Copilot, Cursor, Codex, Gemini, OpenCode. Uses a SQLite mail system for inter-agent communication (8 message types, broadcast addressing). Agent hierarchy: Orchestrator > Coordinator > Supervisor > Workers. Three-tier watchdog system. Tiered merge queue with 4-level conflict resolution.
Explicitly warns: "Agent swarms are not a universal solution."
Best for: Teams using diverse agent runtimes that need merge coordination as a first-class feature.
MyClaude — 2.5k stars
MyClaude | Go, Python, JS
Two-agent model: Claude Code as Orchestrator (planning + verification), codeagent-wrapper as Executor across Codex, Claude, Gemini, and OpenCode backends. 5 workflow modules:
| Module | Purpose |
|---|---|
do | 5-phase feature development |
omo | Multi-agent with routing |
bmad | Enterprise agile with 6 agents |
essentials | 11 core commands |
sparv | Specify-Plan-Act-Review-Vault |
Best for: Developers who want a practical, well-structured workflow system.
類別 3：多 Agent 平台（Agent 不可知）
Ruflo——21.5k 星
Ruflo | MIT | 前身為「Claude Flow」
功能最豐富的編排平台。60+ 個專門 agent 組織成由「queen」領導的 swarm（Strategic、Tactical、Adaptive）。功能：Q-Learning 路由器、8 個 expert 的 Mixture-of-Experts、42+ skill、多種共識演算法（Byzantine、Raft、Gossip、CRDT）、WASM agent booster（簡單轉換 <1ms）、多 provider 支援（Claude、GPT、Gemini、Cohere、Ollama）。宣稱延伸約 250% 的 Claude Code 使用量，token 優化 30-50%。
最適合： 想要最大功能廣度的團隊。
但書： 關於自學習和共識演算法的宣稱需要仔細評估。
Agent Orchestrator（ComposioHQ）——4.5k 星
Agent Orchestrator | MIT | TypeScript
這個領域中架構最乾淨的方案。8 個可替換的 plugin 插槽，每個都由 TypeScript interface 定義：
| 插槽 | 選項 |
|---|---|
| Runtime | tmux / Docker / K8s |
| Agent | Claude Code / Codex / Aider / OpenCode |
| Workspace | worktree / clone |
| Tracker | GitHub / Linear |
| SCM | Git |
| Notifier | Slack / webhook |
| Terminal | tmux |
| Lifecycle | hooks |
Agent 能自主修復 CI 失敗、回應 review 留言、開 PR。3,288 個測試案例。
最適合： 重視乾淨架構和可擴展性的團隊。
Overstory——1k 星
Overstory | MIT | TypeScript/Bun
最不綁定 runtime 的編排器。支援 8 種 agent runtime：Claude Code、Sapling、Pi、Copilot、Cursor、Codex、Gemini、OpenCode。使用 SQLite 郵件系統進行 agent 間通訊（8 種訊息類型、廣播定址）。Agent 層級：Orchestrator > Coordinator > Supervisor > Worker。三層 watchdog 系統。分層 merge queue，4 級衝突解決。
明確警告：「Agent swarm 不是萬能解法。」
最適合： 使用多種 agent runtime、需要 merge 協調作為一級功能的團隊。
MyClaude——2.5k 星
MyClaude | Go、Python、JS
雙 agent 模型：Claude Code 作為 Orchestrator（規劃 + 驗證），codeagent-wrapper 作為 Executor，跨 Codex、Claude、Gemini 和 OpenCode 後端。5 個工作流模組：
| 模組 | 用途 |
|---|---|
do | 5 階段功能開發 |
omo | 多 agent 含路由 |
bmad | 企業敏捷，6 個 agent |
essentials | 11 個核心指令 |
sparv | 指定-計畫-執行-審查-儲存 |
最適合： 想要實用、結構良好的工作流系統的開發者。
Category 4: Session Managers
These tools focus on managing multiple CLI agent sessions — no task decomposition, no dependency graphs, just clean multi-agent management.
cmux — 7.5k stars
cmux | AGPL-3.0
Native macOS terminal app (Swift/AppKit) using Ghostty's GPU-accelerated rendering. Agent notification system with blue rings on panes. Built-in scriptable WebKit browser. Premium terminal emulator with agent-aware features.
Claude Squad — 6.4k stars
Claude Squad | Go | AGPL-3.0
The most popular pure session manager. TUI for managing multiple Claude Code, Codex, Gemini, and Aider instances in isolated tmux sessions with git worktrees. Dead simple: n new session, o attach, s push. Installable via brew install claude-squad.
Best for: Developers who want multi-agent parallelism with zero orchestration overhead.
dmux — 1.2k stars
dmux | MIT
The widest agent support: Claude Code, Codex, OpenCode, Cline CLI, Gemini CLI, Qwen CLI, Amp CLI, Pi CLI, Cursor CLI, Copilot CLI, Crush CLI. Multi-select agent launches, AI-generated branch names, smart merge+cleanup, lifecycle hooks, macOS notifications.
Best for: Developers using many different agent CLIs who want a unified launcher.
Mux (Coder) — 1.4k stars
Mux | By Coder (remote dev environments)
Desktop + browser app for parallel agentic development. Custom agent loop with Plan/Exec mode. Supports Claude, GPT-5, Grok, Ollama, OpenRouter. VS Code extension. SSH-based remote execution.
Best for: Teams already using Coder for remote development environments.
類別 4：Session 管理器
這些工具專注於管理多個 CLI agent session——沒有任務分解、沒有相依圖，就是乾淨的多 agent 管理。
cmux——7.5k 星
cmux | AGPL-3.0
原生 macOS 終端應用程式（Swift/AppKit），使用 Ghostty 的 GPU 加速渲染。Agent 通知系統在面板上顯示藍色光圈。內建可腳本化的 WebKit 瀏覽器。具備 agent 感知功能的進階終端模擬器。
Claude Squad——6.4k 星
Claude Squad | Go | AGPL-3.0
最受歡迎的純 session 管理器。TUI 介面管理多個 Claude Code、Codex、Gemini 和 Aider 實例，在隔離的 tmux session 中搭配 git worktree。超級簡單：n 新 session、o 附加、s 推送。可透過 brew install claude-squad 安裝。
最適合： 想要多 agent 平行性、零編排開銷的開發者。
dmux——1.2k 星
dmux | MIT
支援最多 agent 的方案：Claude Code、Codex、OpenCode、Cline CLI、Gemini CLI、Qwen CLI、Amp CLI、Pi CLI、Cursor CLI、Copilot CLI、Crush CLI。多選 agent 啟動、AI 生成分支名稱、智慧 merge + 清理、lifecycle hook、macOS 通知。
最適合： 使用多種不同 agent CLI、想要統一啟動器的開發者。
Mux（Coder）——1.4k 星
Mux | 來自 Coder（遠端開發環境）
桌面 + 瀏覽器應用程式，用於平行 agentic 開發。自訂 agent 迴圈，Plan/Exec 模式。支援 Claude、GPT-5、Grok、Ollama、OpenRouter。VS Code 擴充套件。SSH 遠端執行。
最適合： 已經使用 Coder 遠端開發環境的團隊。
Category 5: Codex CLI & OpenAI Ecosystem
Symphony (OpenAI) — 13.1k stars
Symphony | Elixir | Engineering Preview
OpenAI's official answer to multi-agent orchestration. Monitors work boards (Linear), spawns autonomous agents, provides proof of work (CI status, complexity analysis, walkthrough videos), and lands PRs when approved. Spec-driven design (SPEC.md) with Elixir reference implementation.
Status: "Low-key engineering preview" — not production-ready but architecturally significant as OpenAI's vision for agentic coding.
Codex Orchestrator — 235 stars
Codex Orchestrator | MIT
Designed to work with Claude Code, not replace it. "Claude handles strategic thinking; Codex handles deep coding work." Delegates tasks to Codex agents via tmux sessions. Plugin pipeline: Ideation > Research > Synthesis > PRD > Implementation > Review > Testing.
Best for: Teams that want Claude Code for planning and Codex for execution.
類別 5：Codex CLI 與 OpenAI 生態系
Symphony（OpenAI）——13.1k 星
Symphony | Elixir | Engineering Preview
OpenAI 對多 agent 編排的官方回應。監控工作看板（Linear）、啟動自主 agent、提供工作證明（CI 狀態、複雜度分析、演練影片），在核准後合併 PR。Spec 驅動設計（SPEC.md），Elixir 參考實作。
狀態：「低調工程預覽」——尚未 production-ready，但作為 OpenAI 對 agentic coding 的願景具有架構意義。
Codex Orchestrator——235 星
Codex Orchestrator | MIT
設計上與 Claude Code 搭配使用，而非取代它。「Claude 處理策略思考；Codex 處理深度 coding 工作。」透過 tmux session 將任務委派給 Codex agent。Plugin 管線：構思 > 研究 > 綜合 > PRD > 實作 > 審查 > 測試。
最適合： 想用 Claude Code 做規劃、Codex 做執行的團隊。
Category 6: Desktop Orchestrators
Aperant — 13.3k stars
Aperant | AGPL-3.0 | Formerly "Auto-Claude"
Electron desktop app. Up to 12 concurrent agent terminals with git worktree isolation, self-validating QA loops, AI-powered conflict resolution. GitHub/GitLab/Linear integration. Memory layer retaining insights across sessions. Requires Claude Pro/Max subscription.
Best for: Developers who prefer GUI over TUI and want visual multi-agent management.
Dorothy — 211 stars
Desktop app with a unique "Super Agent" meta-agent that programmatically controls other agents via MCP tools. Automations poll GitHub/JIRA and spawn agents. Kanban task management with skill-based agent assignment. Scheduled tasks via cron. Telegram/Slack remote control.
Best for: Teams wanting GitHub/JIRA-driven agent automation with remote control.
Jean — 643 stars
Jean | Tauri (Rust)
Tauri desktop app managing Claude CLI, Codex CLI, and OpenCode across projects and worktrees. Integrated terminal, diff viewer, GitHub issue/PR investigation. Magic commands for code review, AI commit messages, PR generation. Built with React 19 + Rust.
Best for: Developers who want a lightweight, native desktop experience (Tauri is much lighter than Electron).
AI Maestro — 532 stars
AI Maestro | Dashboard
Features the Agent Messaging Protocol (AMP) for direct agent-to-agent communication with priority levels and cryptographic signatures. Peer mesh network for multi-machine support without a central server. Three-layer persistent intelligence (memory, code graph, auto-docs). Supports Claude Code, Aider, Cursor, Copilot.
Best for: Distributed teams running agents across multiple machines.
類別 6：桌面編排器
Aperant——13.3k 星
Aperant | AGPL-3.0 | 前身為「Auto-Claude」
Electron 桌面應用程式。最多 12 個並行 agent 終端搭配 git worktree 隔離、自驗證 QA 迴圈、AI 驅動的衝突解決。GitHub/GitLab/Linear 整合。跨 session 的記憶層。需要 Claude Pro/Max 訂閱。
最適合： 偏好 GUI 而非 TUI、想要視覺化多 agent 管理的開發者。
Dorothy——211 星
桌面應用程式，特色是「Super Agent」meta-agent，透過 MCP 工具程式化控制其他 agent。自動化機制輪詢 GitHub/JIRA 並啟動 agent。Kanban 任務管理，基於技能分配 agent。透過 cron 排程任務。Telegram/Slack 遠端控制。
最適合： 想要 GitHub/JIRA 驅動的 agent 自動化搭配遠端控制的團隊。
Jean——643 星
Jean | Tauri（Rust）
Tauri 桌面應用程式，管理 Claude CLI、Codex CLI 和 OpenCode 跨專案與 worktree。整合終端、diff 檢視器、GitHub issue/PR 調查。Magic command 用於 code review、AI commit 訊息、PR 生成。以 React 19 + Rust 建構。
最適合： 想要輕量原生桌面體驗的開發者（Tauri 比 Electron 輕量許多）。
AI Maestro——532 星
AI Maestro | Dashboard
特色是 Agent Messaging Protocol（AMP）——直接 agent 對 agent 通訊，有優先級和加密簽名。Peer mesh 網路支援多機器，不需中央伺服器。三層持久化智慧（記憶、程式碼圖、自動文件）。支援 Claude Code、Aider、Cursor、Copilot。
最適合： 在多台機器上運行 agent 的分散式團隊。
Category 7: Sandboxed Execution
TSK — 150 stars
TSK | Rust
The most security-focused tool in this space. Delegates tasks to agents in Docker containers with domain allowlist enforcement via Squid proxy. Auto-detects project toolchains and builds container images. SQLite-backed task queue with parallel workers. Supports Claude Code and Codex. Task chaining for sequential workflows.
Best for: Security-conscious teams that want agents sandboxed from the host system.
Agent Swarm (desplega-ai) — 267 stars
Lead/worker model. A lead agent receives tasks from Slack/GitHub/email/CLI, breaks them down, and delegates to workers in Docker containers. Compounding memory via embeddings. 9 agent templates. Persistent identity (SOUL.md, IDENTITY.md). Real-time monitoring dashboard.
Best for: Teams that want agents with persistent "personality" and multi-channel task intake.
類別 7：沙箱執行
TSK——150 星
TSK | Rust
這個領域中最注重安全的工具。在 Docker 容器中委派任務給 agent，透過 Squid proxy 強制執行網域白名單。自動偵測專案工具鏈並建構容器映像。SQLite 任務佇列搭配平行 worker。支援 Claude Code 和 Codex。任務串接用於循序工作流。
最適合： 注重安全、想讓 agent 與主機系統隔離的團隊。
Agent Swarm（desplega-ai）——267 星
Lead/Worker 模型。Lead agent 從 Slack/GitHub/email/CLI 接收任務，分解後委派給 Docker 容器中的 worker。透過 embeddings 的累積記憶。9 個 agent 模板。持久化身份（SOUL.md、IDENTITY.md）。即時監控儀表板。
最適合： 想要 agent 有持久化「個性」和多通道任務接收的團隊。
Category 8: Lightweight / Skill-Based
Swarms (am-will) — 144 stars
Swarms | Shell
Pure Shell-based orchestration skills for Claude Code and Codex. Two-phase: swarm-planner creates dependency-ordered task plans, parallel-task launches independent agents in waves. Minimal dependencies — just shell scripts.
Best for: Minimalists who want orchestration without installing anything.
Sudocode — 232 stars
Git-native agent memory system that lives in your repo. 4-tiered abstraction: Specs > Issues > Agents > Artifacts. Dual representation: human-readable Markdown + machine-optimized JSONL/SQLite. Topological ordering for issue execution. IDE integration for VS Code, Cursor, Windsurf, IntelliJ, Zed, Xcode.
Best for: Teams that want agent state to live in the repo, not a separate database.
Claude MPM — 85 stars
47+ specialized agents with PM orchestration for automatic task routing. 44+ bundled skills. Session management with auto-pause at token thresholds. Extensive MCP integrations (Google Workspace, Notion, Confluence, Slack, Jira/Linear/GitHub).
Best for: Solo developers who want a project-manager-in-a-box with broad integrations.
類別 8：輕量 / Skill 型
Swarms（am-will）——144 星
Swarms | Shell
純 Shell 的 Claude Code 和 Codex 編排 skill。兩階段：swarm-planner 建立有相依順序的任務計畫，parallel-task 以波次啟動獨立 agent。最少相依——只有 shell 腳本。
最適合： 不想安裝任何東西就想要編排功能的極簡主義者。
Sudocode——232 星
住在你 repo 裡的 git 原生 agent 記憶系統。4 層抽象：Spec > Issue > Agent > Artifact。雙重表示：人類可讀的 Markdown + 機器最佳化的 JSONL/SQLite。拓撲排序決定 issue 執行順序。IDE 整合支援 VS Code、Cursor、Windsurf、IntelliJ、Zed、Xcode。
最適合： 想讓 agent 狀態住在 repo 裡而非獨立資料庫的團隊。
Claude MPM——85 星
47+ 個專門 agent，PM 編排自動路由任務。44+ 個內建 skill。Session 管理在 token 門檻時自動暫停。廣泛的 MCP 整合（Google Workspace、Notion、Confluence、Slack、Jira/Linear/GitHub）。
最適合： 想要一站式專案經理、且有廣泛整合的單人開發者。
Category 9: Mobile & Remote Access
Happy Coder — 15.5k stars
Happy Coder | TypeScript
Mobile/web client for Claude Code and Codex with real-time voice, end-to-end encryption, and push notifications. Not an orchestrator but enables remote agent supervision from your phone. The happy agent command spawns and monitors remote sessions.
Best for: Developers who want to monitor and interact with agents while away from their desk.
類別 9：行動與遠端存取
Happy Coder——15.5k 星
Happy Coder | TypeScript
Claude Code 和 Codex 的行動/網頁客戶端，支援即時語音、端對端加密和推播通知。不是編排器，但能從手機遠端監督 agent。happy agent 指令啟動並監控遠端 session。
最適合： 離開桌面時仍想監控和互動 agent 的開發者。
Decision Matrix: Which Tool Should You Use?
| Your Situation | Recommended Tool | Why |
|---|---|---|
| I just want parallel Claude Code sessions | Claude Squad | Simplest, zero config, brew install |
| I want parallel sessions with many different agents | dmux | Widest agent support (11+ CLIs) |
| I want full business orchestration with governance | Paperclip | Only tool with board governance, budgets, org charts |
| I want git-native orchestration for a dev team | Gas Town | Deepest git integration, TOML workflows |
| I want clean plugin architecture I can extend | Agent Orchestrator | 8 swappable slots, 3,288 tests |
| I want to supercharge Claude Code with plugins | wshobson/agents | 112 agents, 146 skills, native integration |
| I want smart model routing to save costs | oh-my-claudecode | 30-50% cost reduction via intelligent routing |
| I want maximum features and don't mind complexity | Ruflo | 60+ agents, Q-Learning, WASM, consensus |
| I want sandboxed agent execution | TSK | Docker isolation with network allowlists |
| I want a visual desktop app | Aperant | 12 concurrent terminals, Electron |
| I want lightweight native desktop (not Electron) | Jean | Tauri/Rust, much lighter |
| I want agents on my phone | Happy Coder | Mobile client with push notifications |
| I want zero dependencies, just shell scripts | Swarms | Pure shell, no install |
| I want the official OpenAI approach | Symphony | OpenAI's spec-driven vision |
決策矩陣：你該用哪個工具？
| 你的情況 | 推薦工具 | 原因 |
|---|---|---|
| 我只想要平行的 Claude Code session | Claude Squad | 最簡單，零設定，brew install |
| 我想用很多不同 agent 的平行 session | dmux | 最廣 agent 支援（11+ 種 CLI） |
| 我想要完整的商業編排與治理 | Paperclip | 唯一有董事會治理、預算、組織架構的工具 |
| 我想要 git 原生的開發團隊編排 | Gas Town | 最深的 git 整合，TOML workflow |
| 我想要乾淨的 plugin 架構可以擴展 | Agent Orchestrator | 8 個可替換插槽，3,288 個測試 |
| 我想用 plugin 強化 Claude Code | wshobson/agents | 112 個 agent、146 個 skill、原生整合 |
| 我想要智慧模型路由來省錢 | oh-my-claudecode | 透過智慧路由減少 30-50% 成本 |
| 我想要最多功能，不介意複雜度 | Ruflo | 60+ agent、Q-Learning、WASM、consensus |
| 我想要沙箱化的 agent 執行 | TSK | Docker 隔離搭配網路白名單 |
| 我想要視覺化桌面應用 | Aperant | 12 個並行終端，Electron |
| 我想要輕量原生桌面（非 Electron） | Jean | Tauri/Rust，輕量許多 |
| 我想在手機上操作 agent | Happy Coder | 行動客戶端搭配推播通知 |
| 我想要零相依，只要 shell 腳本 | Swarms | 純 shell，不需安裝 |
| 我想要 OpenAI 官方方案 | Symphony | OpenAI 的 spec 驅動願景 |
The Maturity Warning
Every single project in this list is less than one year old. Most are less than three months old. The entire CLI-based agent orchestration category effectively did not exist before mid-2025.
What this means:
- Expect breaking changes — pin versions, update intentionally
- No security audits — none of these have undergone formal security review
- Documentation is sparse — most projects have a README and little else
- Bus factor is 1 — most are maintained by a single developer or small team
- Stars ≠ maturity — 30k stars in two weeks means hype, not stability
Practical advice:
- Start with Claude Squad or dmux — they're simple, low-risk session managers
- Graduate to Gas Town or Agent Orchestrator when you need real task coordination
- Consider Paperclip only when you need business-level orchestration with governance
- Always keep your agents' work in git worktrees so you can easily discard bad outputs
- Set budget limits wherever the tool supports it — runaway agents are expensive
成熟度警告
這份清單中的每一個專案都不到一年。大多數不到三個月。整個 CLI 型 agent 編排類別在 2025 年中之前基本不存在。
這意味著：
- 預期有破壞性變更 —— 鎖定版本，有意識地更新
- 沒有安全審計 —— 這些專案沒有經過正式的安全審查
- 文件稀疏 —— 大多數專案只有 README，其他很少
- Bus factor 是 1 —— 大多數由單一開發者或小團隊維護
- 星數 ≠ 成熟度 —— 兩週 30k 星意味著熱度，不是穩定性
實用建議：
- 從 Claude Squad 或 dmux 開始——簡單、低風險的 session 管理器
- 當你需要真正的任務協調時，升級到 Gas Town 或 Agent Orchestrator
- 只在你需要有治理的商業層級編排時才考慮 Paperclip
- 永遠讓 agent 的工作在 git worktree 中，這樣你可以輕鬆丟棄不好的產出
- 在工具支援的地方設定預算限制——失控的 agent 很昂貴