ClawTeam Multi-Agent Coordination
Turn multiple AI coding agents into a self-organizing team: a leader splits tasks, spawns workers, coordinates via inboxes, and tracks everything on a kanban board — no human micromanagement.
Data flow is transparent (local ~/.clawteam storage, git worktree isolation, snapshot/restore rollback) and license/upstream attribution (HKUDS fork) is clear; deducted for spawn defaulting to skip-permissions, autonomous multi-agent process spawning and instructions to modify the user's OpenClaw exec-approvals config without explicit confirmation gates — least privilege and confirmation are incomplete; publisher identity unverified.
SKILL.md and both references are highly self-consistent, documenting command options, data models and even failure anti-patterns (task.owner misuse); repo includes CI and tests. But a static review cannot verify key-path reproduction; deducting for unverified install paths, dependence on tmux and multiple external CLI agents, and unvalidated behavior on abnormal input, capped at 10.
Trigger conditions are detailed (explicit keywords plus complexity heuristics), scenarios and non-default agents well covered, with signs of Chinese-provider support (kimi/qwen/moonshot preset); deducted for undeclared mainland-China reachability given reliance on pip ecosystem, overseas CLI agents and services, and limited non-fit boundaries (e.g., behavior without a git repo).
Well-layered structure (SKILL.md plus cli-reference and workflows), with version, license, install and troubleshooting notes; deducted for version mismatch (SKILL.md 0.3.0+openclaw1 vs pyproject +openclaw2), a path contradiction (README installs skills/openclaw/SKILL.md while the assessed path is skills/clawteam/SKILL.md), no changelog, and maintenance responsibility resting solely on an unverified fork author.
The core multi-agent orchestration task is fully described with clear output formats (--, kanban) and plausible marginal value over manual coordination; but under static review there is no verifiable evidence of directly usable representative outputs, and README case results (val_bpb gain, 30 GPU-hours) rely on external links not auditable in this repo — scored at or below the static cap.
The repo contains a CI workflow and a test suite targeting spawn adapters, which is auditable primary material; however tests cover adapter command preparation, not the orchestration key paths the skill documents, and no independent third-party execution evidence exists — capped at 5 and deducted for thin coverage.
- spawn defaults to skip-permissions; agents will execute tasks autonomously. Confirm whether you need --no-skip-permissions or skip_permissions=false before use.
- Install instructions modify the user's OpenClaw exec-approvals config (security=allowlist plus allowlist entry); review whether this matches your security policy.
- The README skill install path (skills/openclaw/SKILL.md) contradicts the assessed path (skills/clawteam/SKILL.md); verify the actual file location in the repo before deploying.
- Core function depends on pip, tmux, multiple overseas CLI agents and external services; mainland-China reachability is undeclared — verify connectivity beforehand.
- Publisher is unverified; version fields disagree between SKILL.md and pyproject; no changelog — update and maintenance path unclear.
What does this skill do, and when should you use it?
ClawTeam is a framework-agnostic CLI tool (clawteam) for coordinating multiple AI agents as a team. This repository is a fork of HKUDS/ClawTeam with OpenClaw as the deeply integrated default agent. The skill (skills/clawteam/SKILL.md, version 0.3.0+openclaw1) teaches agents when and how to use the clawteam commands: creating teams, spawning workers, managing task dependencies, messaging, and monitoring boards. All state lives as JSON files in ~/.clawteam/ with no database or server. The repo is MIT-licensed and bundles 6 skills; this profile covers only the ClawTeam skill.
Drives everything through the clawteam CLI: creates teams (team spawn-team) and spawns worker agents (spawn), each with its own git worktree and identity; maintains a shared task board (pending/in_progress/completed/blocked) with --blocked-by dependency chains, priorities, and auto-unblocking; provides per-agent file-based inboxes for point-to-point and broadcast messaging (receive is destructive, peek is not); builds reusable runtime configs via preset/profile with smoke-testing via profile test; offers git context tools (context conflicts/log/inject) for cross-worktree overlap checks and handoffs; plus snapshots and restore, monitoring dashboards (board show/live/attach/serve/gource), plan approval, and graceful lifecycle management. Every command supports -- output.
- An OpenClaw user wants a large task ('build a full-stack app', 'refactor the codebase') split across parallel agents
- A leader agent needs to autonomously spawn workers, assign tasks along dependency chains, and merge results on completion
- Parallel feature development where each agent must work in an isolated git worktree to avoid merge conflicts
- A team needing a terminal kanban or Web UI to monitor progress, inbox counts, and message history across agents
- Running workers on non-default providers/models (e.g. Moonshot Kimi) configured via profile+preset instead of hardcoded env vars
- Checking for overlapping changes across worktrees before reassigning or merging work (context conflicts)
What are this skill's strengths and limitations?
- Framework-agnostic: any CLI coding agent (openclaw, claude, codex, gemini, kimi, nanobot, opencode, hermes) can participate
- Needs only a filesystem and tmux — no Redis, message queues, or databases; atomic JSON writes in ~/.clawteam/
- Git worktree isolation prevents merge conflicts between parallel agents; branches named clawteam/{team}/{agent}
- Tasks support dependency chains, priorities, auto-unblocking, and blocking task wait
- Strong ops tooling: snapshot/restore, conflict checks, plan approval, graceful shutdown, cost dashboard (v0.3.0)
- All commands support -- machine-readable output for scripted agent use
- Full tmux workflow is Linux/macOS-first; native Windows falls back to subprocess and board attach is unavailable (use board serve or WSL)
- Install path is fiddly: must pip install -e . from this repo, plus PATH setup and OpenClaw exec-approval allowlist configuration
- The upstream PyPI clawteam package and the npm namesake both install the wrong thing — an easy beginner trap
- Hermes workers may complete tasks without sending inbox messages, requiring manual tmux scrollback capture (known upstream issue)
- No automated test suite or independent benchmarks are shown in the repo; the ML research case is quoted from a third party (novix-science/autoresearch)
- spawn defaults to skip-permissions, which you should evaluate for security
How do you install this skill?
The skill lives in the win4r/ClawTeam-OpenClaw repo (6 skills bundled). Important: do NOT pip install clawteam (that's the upstream PyPI version defaulting to claude) and do NOT npm-install it (an unrelated name-squatting package). Correct install:
git clone https://github.com/win4r/ClawTeam-OpenClaw.git
cd ClawTeam-OpenClaw
pip install -e .
Prerequisites: Python 3.10+, tmux (default spawn backend), at least one CLI coding agent (openclaw is the fork default), a git repo. Optional P2P transport: pip install -e ".[p2p]". Symlink the clawteam binary into ~/bin so spawned shells can find it. On Linux/macOS/WSL you can alternatively run bash scripts/install-openclaw.sh. The README gives the skill-copy command for skills/openclaw/SKILL.md; exact placement of this particular skill file is not documented — mirror the same structure into your skills directory.
How do you use this skill?
Basic flow:
export CLAWTEAM_AGENT_ID="leader-001" CLAWTEAM_AGENT_NAME="leader" CLAWTEAM_AGENT_TYPE="leader"
clawteam team spawn-team my-team -d "Project team" -n leader
clawteam spawn --team my-team --agent-name worker1 --task "Implement the auth module"
clawteam inbox send my-team worker1 "Start implementing"
clawteam board live my-team --interval 3
Or simply prompt an agent: "Build a full-stack todo app with auth, database, and React frontend. Use clawteam to split the work." The agent auto-creates a team, tasks, and workers. Workers follow a loop: task list to check assignments, inbox receive for new instructions, lifecycle idle when idle, until the leader explicitly approves shutdown. Trigger phrases include create a team, spawn agents, check team status, view kanban board.
How does this skill compare with similar options?
The README contrasts ClawTeam with other multi-agent frameworks: it is driven by the AI agents themselves (not humans writing orchestration code), uses only filesystem+tmux rather than Redis/message queues/databases, supports any CLI agent, and isolates via real git worktrees. The upstream is HKUDS/ClawTeam; this fork differentiates with the default OpenClaw agent, per-agent session isolation, and OpenClaw-specific adaptations.