ClawTeam-OpenClaw

ClawTeam-OpenClaw is a fork of HKUDS/ClawTeam that makes OpenClaw the default agent for multi-agent swarm coordination. You state a goal; a leader agent uses the clawteam CLI to create a team, spawn worker agents, split the work into tasks with dependencies, exchange messages through inboxes, and merge the results. Each worker runs in its own git worktree (branch clawteam/{team}/{agent}) inside a tmux window or subprocess, with a coordination prompt injected automatically.

Workers can be OpenClaw (default), Claude Code, Codex, nanobot, Hermes Agent, Cursor (experimental), or any custom CLI script. The fork adds per-agent session isolation, automatic exec-approval configuration for OpenClaw, and hardened spawn backends, and syncs upstream fixes. Skill files teach OpenClaw and Hermes agents to use ClawTeam from natural-language requests such as "Create a 5-agent team to build a web app".

All state lives as JSON files under ~/.clawteam/ (teams, tasks, inboxes, workspaces) with atomic writes and file locking; no database or server is required. Messaging uses a file transport by default or ZeroMQ peer-to-peer. Linux and macOS use tmux; Windows 10/11 falls back to a subprocess backend, with board attach requiring tmux and board serve available as the Web UI alternative. Requires Python 3.10 or later.

Features

• Agent-driven orchestration: the leader agent spawns, assigns, monitors, and merges without human-written orchestration code
• Multi-agent backends: OpenClaw, Claude Code, Codex, nanobot, Hermes Agent, Cursor, and custom scripts as spawn targets
• Worktree isolation: every agent gets its own git worktree, with checkpoint, merge, and cleanup commands
• Task dependencies: shared kanban with pending, in_progress, completed, and blocked states, --blocked-by chains, and auto-unblock
• Inter-agent messaging: point-to-point inboxes and broadcast over file or ZeroMQ transport
• Monitoring: board show, board live, tiled tmux board attach, a real-time board serve Web UI, and a board cost token and cost dashboard
• Team templates: TOML files (hedge-fund, research-paper, code-review, strategy-room) launched with one command and variable substitution
• Resilience controls: circuit breaker states, retry with backoff, idempotency keys, spawn validation with rollback
• Per-agent model resolution: a priority chain lets one team mix Claude, GPT, and Qwen models
• Runtime injection: runtime inject/state/watch sends messages to running agents; --json output on all commands