Shoal — AI skill for Claude Code
只需定義一次 agent role,shoal 就能依照個廠牌 agent 配好模型、權限與 prompt,自動 orchestration;毋需撰寫任何程式.
How to install Shoal
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open miyago9267/shoal and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Shoal does
只需定義一次 agent role,shoal 就能依照個廠牌 agent 配好模型、權限與 prompt,自動 orchestration;毋需撰寫任何程式。 / Define agent roles once. shoal resolves the model, permissions and prompt for each agent tool; adding a new tool is configuration, not code.
Alternatives in AI
- Open Multi Agent — TypeScript multi-agent orchestration engine — one runTeam() call from goal to result 5.8k ★
- Define — Definition phase - Clarify and scope problems with multi-AI consensus 2.8k ★
- Loushang — AI-native agent harness for coding workflows by python: multi-model LLM orchestration, stateful sessions, tool 1.3k ★
README
shoal
只需定義一次 agent role,shoal 就能依照各廠牌 agent 配好模型、權限與 prompt,自動 orchestration;毋需撰寫任何程式。
繁體中文 · [English](docs/README.en.md) · [简体中文](docs/README.zh-CN.md)
它解決什麼
- 一份 role 定義,五個 host 各自拿到合適的模型、權限和 prompt。
- 主 session 負責判斷與整合,改檔和長指令派給對應的 role。
- 派工規則由 guard 在每個 host 強制執行,不靠模型自覺。
- 要支援新工具,填設定就好,不用寫程式。
運作方式

role 寫在 core,每個 host 用 binding 說明自己的模型與權限,再由 `render.py` 產出安裝檔。

role 只寫 tier,不寫模型名稱;每個 host 從自己的模型目錄挑出符合的那一個。

主 session 先看工作的形狀和風險,再決定自己做、派給哪個 role,或走完整的審查流程。

主 session 沒派可寫 role 就改檔,或 subagent 不是 shoal 的可寫 role 卻改檔,都會被擋;出口是改派可寫的 role,或在 prompt 加 `#direct`。
快速開始
# Codex CLI
bash install/install.sh --codex-home "${CODEX_HOME:-$HOME/.codex}"
# Claude Code(plugin,裝最新穩定版)
claude plugin marketplace add miyago9267/shoal
claude plugin install shoal@shoal
# Grok Build(不加 --apply 是 dry-run)
python3 tools/install_grok.py
# OpenCode(全域)
sh hosts/opencode/plugin/install/install.sh --global --enable
Gemini/agy 與不用 plugin 的 Claude Code 見 [INSTALL.md](./INSTALL.md)。
模組
只裝需要的部分:`python3 tools/shoal_install.py --host ` 預設只印出會執行的指令,加 `--apply` 才安裝。
不要派工守門就加 `--without dispatch-guard`,`--list` 列出模組與已安裝狀態。
外部 guard 經 guard chain 接入:`--chain-add` 登記,不需要改 shoal。
細節見 [INSTALL.md 的模組](./INSTALL.md#模組)。
支援的 host
| Host | 安裝方式 | guard |
|---|---|---|
| Claude Code | plugin,或 tools/install_hooks.py |
hook,預設擋 |
| Codex CLI | install/install.sh |
hook,需在 /hooks 核准一次 |
| Gemini/agy | hosts/agy/dist 加 tools/install_hooks.py |
hook,預設擋,沒有 #direct |
| Grok Build | tools/install_grok.py |
原生 hooks |
| OpenCode | install.sh --enable |
plugin 內建 |
快速與嚴謹
預設走快速路徑:讀檔、診斷和整合由主 session 直接做,改檔派給可寫的 role。遇到風險或大型工作才走嚴謹路徑:計畫審查、核准,完成後再驗證。想讓主 sess
Related Skills
AI Execution Router
Claude Code skill that routes software-engineering issues to an execution profile (model role, effort, effort
Omp Agent Role Router
omp extension: routes Claude Code / Codex plugin subagents to omp model roles by declared tier
Modelweaver
Multi-provider model orchestration proxy for Claude Code. Route agent roles (planning, coding, research) to di
Pilotfish Grok
Grok Build–native multi-model orchestration, inspired by pilotfish. Frontier main session plans; role agents e
Plenipo
Free, local-first Windows desktop app for AI workforce and multi-agent orchestration. Coordinate Claude Code,
Hawk Hooks
One registry for all your AI coding tools. Define skills, hooks, agents, prompts, and MCP servers once; hawk s
Related Agents
Courtroom Juror
Courtroom role — one specialist seat on the jury; rules on the motion strictly from the specialty named in the
Policy Agent
Creates secure Pundit authorization policies with comprehensive RSpec tests and scope restrictions. Use when a
Sc Recon Expander
Task-less smart-contract recon expander — resolves proxies/diamonds/roles/linked addresses per chain and retur