clawteam

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

clawteam — Agent Swarm Intelligence

clawteam — Agent集群智能

Keyword:
clawteam
·
agent swarm
·
spawn agents
·
multi-agent team
One command → a self-organizing team of AI agents that spawn workers, coordinate tasks, isolate workspaces with git worktrees, and deliver results autonomously.

"Agents spawn agents, delegate tasks, and deliver results."

关键词:
clawteam
·
agent swarm
·
spawn agents
·
multi-agent team
一条命令 → 由AI Agent组成的自组织团队，可生成工作Agent、协调任务、通过git worktree隔离工作区，并自主交付结果。

"Agent生成Agent、分发任务并交付结果。"

When to use this skill

适用场景

Spawn parallel AI agent workers for large, decomposable tasks
Run autonomous ML experiments across multiple GPU workers
Orchestrate full-stack development with specialized agents (API, backend, frontend, tests)
Run investment analysis with multi-specialist agent teams
Coordinate agents across Claude Code, Codex, OpenClaw, nanobot, or Cursor
Monitor live agent progress via tmux tiled views or web Kanban dashboard
Track task dependencies and automatically unblock agents when prerequisites complete

为大型可拆解任务生成并行AI Agent工作节点
跨多个GPU工作节点运行自主ML实验
搭配专用Agent（API、后端、前端、测试）编排全栈开发流程
由多领域专家Agent团队运行投资分析
跨Claude Code、Codex、OpenClaw、nanobot、Cursor协调Agent工作
通过tmux平铺视图或网页看板仪表盘实时监控Agent进度
追踪任务依赖，当前置条件完成时自动为Agent解除阻塞

Installation

安装

bash

undefined

bash

undefined

Requires Python 3.10+, tmux, and at least one compatible CLI coding agent

pip install clawteam

Verify installation

clawteam --version clawteam doctor # check tmux, agent CLI, environment


Compatible agents: `claude`, `codex`, `openclaw`, `nanobot`, `cursor`, custom scripts.

---

clawteam --version clawteam doctor # check tmux, agent CLI, environment


兼容Agent: `claude`, `codex`, `openclaw`, `nanobot`, `cursor`, 自定义脚本。

---

Core Concepts

核心概念

Concept	Description
Leader	The orchestrating agent that spawns workers and delegates tasks
Worker	A sub-agent spawned into an isolated git worktree + tmux pane
Team	Named group of workers coordinating on a shared task set
Task	A unit of work with status: `pending` → `in_progress` → `completed` / `blocked`
Inbox	File-based or ZeroMQ P2P message queue per agent
Worktree	Git worktree at branch `clawteam/{team}/{agent}` — isolated file system

概念	描述
主控Agent（Leader）	负责编排工作的Agent，用于生成工作节点并分发任务
工作Agent（Worker）	被生成的子Agent，运行在隔离的git worktree + tmux窗格中
团队（Team）	协同完成同一组任务的工作节点的命名组
任务（Task）	工作单元，状态流转： `pending` → `in_progress` → `completed` / `blocked`
收件箱（Inbox）	每个Agent对应的基于文件或ZeroMQ P2P的消息队列
工作树（Worktree）	分支为 `clawteam/{team}/{agent}` 的Git工作树，即隔离的文件系统

Quick Start

快速开始

Spawn a team from your agent session

从你的Agent会话生成一个团队

bash

undefined

bash

undefined

Leader spawns 3 Claude workers on a task

clawteam spawn claude "Build the REST API" --workers 3 --team api-build

Spawn specialized workers

clawteam spawn claude "Implement auth endpoints" --team api-build --name auth-worker clawteam spawn claude "Implement user CRUD" --team api-build --name crud-worker clawteam spawn codex "Write integration tests" --team api-build --name test-worker

undefined

undefined

Check team status

查看团队状态

bash

clawteam status                      # all teams
clawteam status --team api-build     # specific team
clawteam board --team api-build      # live web Kanban dashboard

bash

clawteam status                      # all teams
clawteam status --team api-build     # specific team
clawteam board --team api-build      # live web Kanban dashboard

Send messages between agents

Agent间发送消息

bash

undefined

bash

undefined

Leader sends task to a worker

clawteam send --to crud-worker "Priority: implement DELETE /users/{id} first"

Worker broadcasts status

clawteam broadcast --team api-build "Auth module complete, unblocking crud-worker"

undefined

clawteam broadcast --team api-build "Auth module complete, unblocking crud-worker"

undefined

Task management

任务管理

bash

clawteam task add "Implement rate limiting" --team api-build
clawteam task list --team api-build
clawteam task done --id <task-id> --team api-build
clawteam task block --id <task-id> --blocked-by <other-id>

bash

clawteam task add "Implement rate limiting" --team api-build
clawteam task list --team api-build
clawteam task done --id <task-id> --team api-build
clawteam task block --id <task-id> --blocked-by <other-id>

Instructions

使用说明

Use this sequence when the user asks to create or run a ClawTeam workflow.

Verify prerequisites first.
- Check Python 3.10+,
```
tmux
```
  ,
```
git
```
  , and at least one supported agent CLI.
- If setup is unclear, run
```
bash scripts/install.sh
```
  or show the minimal install commands.
Pick the team shape.
- Use
```
clawteam template ...
```
  for known patterns like
```
fullstack
```
  ,
```
ml-research
```
  , or
```
hedge-fund
```
  .
- Use explicit
```
clawteam spawn ... --name ...
```
  commands when worker roles need to be custom.
Make worker ownership explicit.
- Give each worker a semantic role such as
```
architect
```
  ,
```
backend
```
  ,
```
tester
```
  , or
```
risk-analyst
```
  .
- If tasks depend on one another, show
```
clawteam task block --id ... --blocked-by ...
```
  .

Show how to monitor and communicate.

Include
```
clawteam status --team <name>
```
.

Include either

clawteam board --team <name>

clawteam monitor --team <name>

Include
```
clawteam send
```
or
```
clawteam broadcast
```
when coordination matters.

Close with the lifecycle.
- Explain how to mark tasks done, inspect costs when relevant, and tear down the team with
```
clawteam teardown --team <name>
```
  .

If the user asks for a recommendation, prefer the smallest team that can execute in parallel without overlapping ownership.

当用户要求创建或运行ClawTeam工作流时，请遵循以下流程：

首先验证前置条件。
- 检查Python 3.10+、
```
tmux
```
  、
```
git
```
  ，以及至少一个支持的Agent CLI是否可用。
- 如果设置不清晰，运行
```
bash scripts/install.sh
```
  或展示最小安装命令。
选择团队架构。
- 已知场景可使用
```
clawteam template ...
```
  ，比如
```
fullstack
```
  、
```
ml-research
```
  、
```
hedge-fund
```
  模板。
- 当需要自定义工作节点角色时，使用显式的
```
clawteam spawn ... --name ...
```
  命令。
明确工作节点权责。
- 为每个工作节点赋予语义化角色，比如
```
architect
```
  、
```
backend
```
  、
```
tester
```
  、
```
risk-analyst
```
  。
- 如果任务间存在依赖，展示
```
clawteam task block --id ... --blocked-by ...
```
  的使用方法。

展示监控与通信方式。

包含
```
clawteam status --team <name>
```
命令。

包含

clawteam board --team <name>

或

clawteam monitor --team <name>

命令。

当协调需求较高时，包含
```
clawteam send
```
或
```
clawteam broadcast
```
命令。

收尾说明生命周期。
- 说明如何标记任务完成、按需查看成本，以及使用
```
clawteam teardown --team <name>
```
  销毁团队。

如果用户需要推荐方案，优先选择可并行执行且权责不重叠的最小规模团队。

Architecture

架构

Leader Agent
    │
    ├── clawteam spawn → Worker 1 (tmux pane, git worktree: clawteam/my-team/worker-1)
    ├── clawteam spawn → Worker 2 (tmux pane, git worktree: clawteam/my-team/worker-2)
    └── clawteam spawn → Worker 3 (tmux pane, git worktree: clawteam/my-team/worker-3)
                │
    ┌───────────┼───────────┐
    │           │           │
 Task 1      Task 2      Task 3
 (independent) (blocked  (blocked
               by T1)     by T2)

State: ~/.clawteam/{team}/
  tasks.json    — task list with deps
  inbox/        — per-agent message queues
  snapshots/    — periodic state snapshots
  costs.json    — token/API cost tracking

Workspace isolation: Each worker operates in

clawteam/{team}/{worker}

git branch → no merge conflicts during parallel development.

Transport: FileTransport (default, atomic tmp+rename) or P2PTransport (ZeroMQ TCP with filesystem fallback).

Leader Agent
    │
    ├── clawteam spawn → Worker 1 (tmux pane, git worktree: clawteam/my-team/worker-1)
    ├── clawteam spawn → Worker 2 (tmux pane, git worktree: clawteam/my-team/worker-2)
    └── clawteam spawn → Worker 3 (tmux pane, git worktree: clawteam/my-team/worker-3)
                │
    ┌───────────┼───────────┐
    │           │           │
 Task 1      Task 2      Task 3
 (independent) (blocked  (blocked
               by T1)     by T2)

State: ~/.clawteam/{team}/
  tasks.json    — task list with deps
  inbox/        — per-agent message queues
  snapshots/    — periodic state snapshots
  costs.json    — token/API cost tracking

工作区隔离: 每个工作节点运行在

clawteam/{team}/{worker}

Git分支上 → 并行开发过程中不会出现合并冲突。

传输层: FileTransport（默认，原子tmp+重命名）或P2PTransport（ZeroMQ TCP，带文件系统降级方案）。

Pre-built Team Templates

预置团队模板

ML Research Team

机器学习研究团队

bash

undefined

bash

undefined

8 agents × H100 GPU experiments

clawteam template ml-research
--agents 8
--task "Optimize transformer attention with val_bpb target 0.97"
--team ml-exp

undefined

clawteam template ml-research
--agents 8
--task "Optimize transformer attention with val_bpb target 0.97"
--team ml-exp

undefined

Full-Stack Development Team

全栈开发团队

bash

clawteam template fullstack \
  --agents 4 \
  --task "Build e-commerce product catalog API" \
  --team shop-api

bash

clawteam template fullstack \
  --agents 4 \
  --task "Build e-commerce product catalog API" \
  --team shop-api

Spawns: api-designer, backend-dev, frontend-dev, test-engineer

undefined

undefined

Investment Analysis Team

投资分析团队

bash

clawteam template hedge-fund \
  --agents 7 \
  --ticker AAPL \
  --team aapl-analysis

bash

clawteam template hedge-fund \
  --agents 7 \
  --ticker AAPL \
  --team aapl-analysis

Spawns: growth-analyst, technical-analyst, fundamental-analyst,

sentiment-analyst, risk-analyst, macro-analyst, synthesizer

---

---

Monitoring

监控

bash

undefined

bash

undefined

Real-time tiled tmux view (split panes per worker)

clawteam monitor --team api-build

Web Kanban dashboard (localhost:7420)

clawteam board --team api-build --port 7420

Show agent costs and token usage

clawteam costs --team api-build

---

clawteam costs --team api-build

---

Configuration

配置

Config file:

~/.clawteam/config.toml

toml

[transport]
type = "file"          # "file" (default) or "p2p" (ZeroMQ)

[workspace]
base_dir = "~/.clawteam"
worktree_pattern = "clawteam/{team}/{agent}"

[spawn]
default_agent = "claude"
tmux_session_prefix = "ct-"

[costs]
budget_limit = 10.0    # USD, optional
alert_threshold = 0.8  # warn at 80% of budget

Environment variables override config:

bash

CLAWTEAM_TRANSPORT=p2p   # use ZeroMQ
CLAWTEAM_BASE_DIR=/tmp/clawteam

配置文件路径:

~/.clawteam/config.toml

toml

[transport]
type = "file"          # "file" (default) or "p2p" (ZeroMQ)

[workspace]
base_dir = "~/.clawteam"
worktree_pattern = "clawteam/{team}/{agent}"

[spawn]
default_agent = "claude"
tmux_session_prefix = "ct-"

[costs]
budget_limit = 10.0    # USD, optional
alert_threshold = 0.8  # warn at 80% of budget

环境变量可覆盖配置:

bash

CLAWTEAM_TRANSPORT=p2p   # use ZeroMQ
CLAWTEAM_BASE_DIR=/tmp/clawteam

CLI Reference

CLI参考

Command	Description
`clawteam spawn <agent> <task>`	Spawn a worker agent with a task
`clawteam status [--team NAME]`	Show team and task status
`clawteam board [--team NAME]`	Open web Kanban dashboard
`clawteam monitor [--team NAME]`	Live tmux tiled view
`clawteam send --to <worker> <msg>`	Send message to a specific worker
`clawteam broadcast --team <name> <msg>`	Broadcast to all team workers
`clawteam task add/list/done/block`	Manage tasks
`clawteam costs [--team NAME]`	Show token/cost usage
`clawteam doctor`	Check environment and dependencies
`clawteam template <name>`	Use pre-built team templates
`clawteam teardown --team NAME`	Shut down team and clean up

命令	描述
`clawteam spawn <agent> <task>`	生成带指定任务的工作Agent
`clawteam status [--team NAME]`	展示团队和任务状态
`clawteam board [--team NAME]`	打开网页看板仪表盘
`clawteam monitor [--team NAME]`	实时tmux平铺视图
`clawteam send --to <worker> <msg>`	向指定工作Agent发送消息
`clawteam broadcast --team <name> <msg>`	向团队所有工作Agent广播消息
`clawteam task add/list/done/block`	管理任务
`clawteam costs [--team NAME]`	展示token/成本消耗
`clawteam doctor`	检查环境和依赖
`clawteam template <name>`	使用预置团队模板
`clawteam teardown --team NAME`	关闭团队并清理资源

Worker Auto-Injection

工作Agent自动注入

When ClawTeam spawns a worker, it injects a coordination prompt teaching the agent to:

Check assigned tasks:

clawteam task list --team {name} --owner {me}

Mark task in progress:
```
clawteam task start --id {id}
```
Report completion:
```
clawteam task done --id {id}
```

Check inbox:

clawteam inbox --team {name} --worker {me}

Send status:

clawteam send --to leader "Task complete: {summary}"

Workers operate autonomously — no human intervention needed once spawned.

当ClawTeam生成工作Agent时，会自动注入协调提示词，指导Agent完成以下操作：

查看分配的任务:

clawteam task list --team {name} --owner {me}

标记任务进行中:
```
clawteam task start --id {id}
```
上报任务完成:
```
clawteam task done --id {id}
```

查看收件箱:

clawteam inbox --team {name} --worker {me}

发送状态:

clawteam send --to leader "Task complete: {summary}"

工作Agent可自主运行 —— 生成后无需人工干预。

Examples

示例

Example 1: Spawn a 3-worker API team

示例1: 生成3个工作节点的API开发团队

bash

clawteam spawn claude "Design the API contracts" --team api-build --name architect
clawteam spawn claude "Implement the REST endpoints" --team api-build --name backend
clawteam spawn codex "Write integration tests" --team api-build --name tester

clawteam task block --id backend-impl --blocked-by api-spec
clawteam task block --id integration-tests --blocked-by backend-impl

clawteam status --team api-build
clawteam board --team api-build

bash

clawteam spawn claude "Design the API contracts" --team api-build --name architect
clawteam spawn claude "Implement the REST endpoints" --team api-build --name backend
clawteam spawn codex "Write integration tests" --team api-build --name tester

clawteam task block --id backend-impl --blocked-by api-spec
clawteam task block --id integration-tests --blocked-by backend-impl

clawteam status --team api-build
clawteam board --team api-build

Example 2: Run parallel ML experiments

示例2: 运行并行ML实验

bash

clawteam template ml-research \
  --agents 4 \
  --task "Sweep learning rate and context length for val_bpb improvement" \
  --team ml-sweep

clawteam monitor --team ml-sweep
clawteam costs --team ml-sweep

bash

clawteam template ml-research \
  --agents 4 \
  --task "Sweep learning rate and context length for val_bpb improvement" \
  --team ml-sweep

clawteam monitor --team ml-sweep
clawteam costs --team ml-sweep

Example 3: Check what a team is doing right now

示例3: 查看团队当前运行状态

bash

clawteam status --team shop-api
clawteam board --team shop-api --port 7420
clawteam inbox --team shop-api --worker backend

bash

clawteam status --team shop-api
clawteam board --team shop-api --port 7420
clawteam inbox --team shop-api --worker backend

Best practices

最佳实践

Name workers semantically —
```
auth-worker
```
,
```
test-engineer
```
, not
```
worker-1
```
Use task dependencies —
```
clawteam task block
```
prevents out-of-order execution
Set budget limits — use
```
budget_limit
```
in config to prevent runaway costs
Use templates for known patterns —
```
ml-research
```
,
```
fullstack
```
,
```
hedge-fund
```
FileTransport for most cases — P2PTransport only needed for cross-machine teams
Monitor with
clawteam board
— web dashboard gives clearest progress view

为工作节点起语义化名称 —— 比如
```
auth-worker
```
、
```
test-engineer
```
，不要用
```
worker-1
```
这类名称
使用任务依赖 ——
```
clawteam task block
```
可避免任务乱序执行
设置预算上限 —— 在配置中使用
```
budget_limit
```
避免超额成本
已知场景使用模板 ——
```
ml-research
```
、
```
fullstack
```
、
```
hedge-fund
```
绝大多数场景使用FileTransport —— 仅跨机器团队需要使用P2PTransport
使用
clawteam board
监控 —— 网页仪表盘提供最清晰的进度视图

References

参考资料

Architecture Overview — transport, workspace isolation, state model
Agent Types & Compatibility — claude, codex, openclaw, nanobot, cursor
Use Cases & Templates — ML research, fullstack dev, investment analysis
ClawTeam GitHub — MIT License
PyPI —
```
pip install clawteam
```

架构概览 — 传输层、工作区隔离、状态模型
Agent类型与兼容性 — claude, codex, openclaw, nanobot, cursor
使用场景与模板 — 机器学习研究、全栈开发、投资分析
ClawTeam GitHub — MIT License
PyPI —
```
pip install clawteam
```