ClawTeam Multi-Agent Swarm Coordination
Turn CLI coding agents into self-organizing teams that split tasks, work in parallel, manage dependencies, message each other, and merge results — one command to launch.
Docs transparently disclose data locations (~/.clawteam), spawn safety (pre-validation, rollback) and exec-approvals config; but the skill auto-confirms workspace trust prompts, recommends allowlist reconfiguration and uses --force cleanup with broad external effects and no per-action user confirmation — deducted.
SKILL.md command tables are self-consistent with troubleshooting and rollback notes; CI and adapter tests exist, but tests do not cover this skill's key paths (team lifecycle, messaging, merging); failure-feedback quality is asserted only — unverified in static review, deducted.
Clear trigger phrases, scenarios and boundaries (tmux dependency, Windows fallback); multilingual README helps Chinese users; but core dependence on OpenClaw/external CLIs and a GitHub install path, with mainland-China reachability undeclared and non-fit ranges thinly disclosed, deducted.
Well-layered docs, clear versioning (0.3.0+openclaw2), MIT license, explicit upstream attribution; but no changelog, maintenance/update path only implicit in the fork claim, and the npm name-squatting warning reveals naming risk, deducted.
The orchestration value proposition is concrete with a complete workflow, but all effect claims (e.g. the 8-GPU research case via an external repo) are statically unverifiable, with no representative usable-output evidence and thin comparative-benefit data, deducted.
CI workflow plus committed tests (good adapter coverage) exist, but tests do not directly exercise this skill's key paths; effectiveness claims lean on external links and marketing-style narrative without strict fact/inference separation, deducted.
- The skill auto-confirms workspace trust prompts and suggests rewriting exec-approvals config; manually review allowlist changes before installing.
- clawteam team cleanup --force irreversibly deletes all team data; save results before use.
- No tests found covering the openclaw skill path itself; rehearse on a small team first.
- Depends on OpenClaw and other external services plus a GitHub install path; mainland-China reachability is undeclared — verify independently.
- Publisher is unverified and an npm name-squatting package exists; install from the local repository per the README only.
What does this skill do, and when should you use it?
ClawTeam is a CLI tool (clawteam) that orchestrates multiple AI coding agents as self-organizing swarms. It isolates each agent in its own git worktree, hosts sessions in tmux windows, and routes communication through filesystem-based inboxes. This skill (skills/openclaw/SKILL.md) teaches an OpenClaw agent to create teams, spawn workers, manage task dependencies, and monitor progress via natural language. The repository is a fork of HKUDS/ClawTeam deeply adapted to make OpenClaw the default agent backend. Three pre-built team templates ship out of the box: hedge-fund, code-review, and research-paper.
When the user expresses intent like 'build a team', 'spawn agents', or 'swarm', the skill guides the agent to: create a team with a leader via clawteam team spawn-team; build dependency chains with clawteam task create --blocked-by (dependents auto-unblock when blockers complete); give each worker its own tmux window plus a git worktree on branch clawteam/{team}/{agent} with an auto-injected coordination prompt via clawteam spawn; exchange point-to-point and broadcast messages via inbox send/broadcast/receive; monitor via board show/attach/serve (terminal kanban, tiled tmux view, or Web dashboard); track spend with cost report/show/budget; merge branches with workspace merge and tear down with team cleanup. All state lives as JSON files under ~/.clawteam/ with fcntl file locking for concurrent safety, and every command supports -- output.
- Full-stack development: say 'create a team to build a web app with auth' — the leader splits work into a dependency chain (API design → backend/frontend → tests), 5 agents develop in parallel in separate worktrees, then everything merges to main
- Investment analysis: clawteam launch hedge-fund boots a 7-agent team (5 parallel analysts, a risk manager, and a portfolio manager) from a single TOML template
- Code review: launch the code-review template so multiple agents divide up review work and consolidate findings
- Research writing: launch the research-paper template to organize literature survey and writing assignments
- Large-scale experimentation: a leader spawns multiple research agents in isolated worktrees and cross-pollinates best configs (README cites an autoresearch run: 2,430 experiments in ~30 GPU-hours)
What are this skill's strengths and limitations?
- Agent self-organization: coordination logic runs in the leader agent — no human-written orchestration code
- Real isolation: each agent works on its own git worktree branch, avoiding parallel-edit conflicts, with merge support on completion
- Zero infrastructure: filesystem + tmux only; state is JSON under ~/.clawteam/, no database or message queue needed
- Automatic dependency resolution, task locking against double-claims, and auto-release of stale locks from dead agents
- Multiple monitoring options: terminal kanban, tiled tmux view, Web dashboard, and -- output on all commands
- Spawn safety: pre-validation of commands, automatic rollback on failed spawns, auto-confirmed trust prompts for Claude Code/Codex worktrees
- Installation is multi-step and manual: PATH symlink, exec-approval configuration, and skill file copying
- Strong local-environment dependency: Python 3.10+, git, tmux (Linux/macOS) and at least one CLI agent — not usable in pure cloud-API setups as-is
- board attach requires tmux and is unavailable on native Windows (README suggests board serve or WSL instead)
- No automated test suite or benchmark data appears in SKILL.md or README; some claimed results (e.g., the 8-GPU experiment) are external case studies cited from elsewhere, not verified in this repo
- Switching the spawn backend away from openclaw (e.g., to claude) hits interactive permission prompts unless extra configuration is done
How do you install this skill?
Prerequisites: Python 3.10+, git, tmux on Linux/macOS, and at least one CLI coding agent (OpenClaw by default). Do NOT pip install clawteam (that's the upstream version defaulting to claude, without OpenClaw adaptations) and do NOT npm install -g clawteam (an unrelated name-squatting package). Correct install:
- git clone https://github.com/win4r/ClawTeam-OpenClaw.git && cd ClawTeam-OpenClaw
- pip install -e . (required — installs from the local repo, not PyPI)
- Ensure clawteam is on PATH: mkdir -p ~/bin && ln -sf "$(which clawteam)" ~/bin/clawteam
- Install this skill: mkdir -p ~/.openclaw/workspace/skills/clawteam && cp skills/openclaw/SKILL.md ~/.openclaw/workspace/skills/clawteam/SKILL.md
- Set OpenClaw exec approvals to 'allowlist' and add the absolute clawteam path to the allowlist (scripts included in the skill). Alternatively run bash scripts/install-openclaw.sh to automate.
How do you use this skill?
After installation, just talk to your OpenClaw agent in natural language — trigger phrases include team, swarm, multi-agent, spawn agents. Example: 'Create a 5-agent team to build a web app' — the agent creates the team, tasks, and workers and starts a background monitoring loop automatically. Manual route: clawteam launch hedge-fund --team fund1 for one-command template launch, or step through team spawn-team → task create → spawn → board serve. Important: keep the default openclaw spawn command; overriding to claude will get agents stuck on interactive permission prompts.
How does this skill compare with similar options?
The README contrasts ClawTeam with other multi-agent frameworks: those require Docker, cloud APIs, YAML configs, Redis/message queues, are driven by humans writing orchestration code, and are framework-agent-specific. ClawTeam needs only pip install plus one prompt, uses filesystem + tmux as infrastructure, accepts any CLI agent, and isolates via real git worktrees. The upstream HKUDS/ClawTeam defaults to claude and lacks OpenClaw adaptations; this fork is the deeply OpenClaw-integrated version.