[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-74242":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":16,"subscribersCount":16,"size":16,"stars1d":14,"stars7d":17,"stars30d":18,"stars90d":16,"forks30d":16,"starsTrendScore":19,"compositeScore":20,"rankGlobal":10,"rankLanguage":10,"license":21,"archived":22,"fork":22,"defaultBranch":23,"hasWiki":24,"hasPages":22,"topics":25,"createdAt":10,"pushedAt":10,"updatedAt":35,"readmeContent":36,"aiSummary":37,"trendingCount":16,"starSnapshotCount":16,"syncStatus":15,"lastSyncTime":38,"discoverSource":39},74242,"ClawTeam-OpenClaw","win4r\u002FClawTeam-OpenClaw","win4r","ClawTeam fork fully adapted for OpenClaw — multi-agent swarm coordination with OpenClaw as the default agent","https:\u002F\u002Fyoutu.be\u002FaZT9d8qrirY",null,"Python",1398,314,9,2,0,13,37,27,20.49,"MIT License",false,"main",true,[26,27,28,29,30,31,32,33,34],"ai-agents","clawdbot","openclaw","openclaw-extension","openclaw-plugin","openclaw-skills","swarm","swarm-intelligence","swarms","2026-06-12 02:03:24","\u003Cp align=\"center\">\n  \u003Ca href=\"README.md\">English\u003C\u002Fa> |\n  \u003Ca href=\"README_CN.md\">简体中文\u003C\u002Fa> |\n  \u003Ca href=\"README_TW.md\">繁體中文\u003C\u002Fa> |\n  \u003Ca href=\"README_JA.md\">日本語\u003C\u002Fa> |\n  \u003Ca href=\"README_KO.md\">한국어\u003C\u002Fa> |\n  \u003Ca href=\"README_FR.md\">Français\u003C\u002Fa> |\n  \u003Ca href=\"README_ES.md\">Español\u003C\u002Fa> |\n  \u003Ca href=\"README_DE.md\">Deutsch\u003C\u002Fa> |\n  \u003Ca href=\"README_IT.md\">Italiano\u003C\u002Fa> |\n  \u003Ca href=\"README_RU.md\">Русский\u003C\u002Fa> |\n  \u003Ca href=\"README_PT-BR.md\">Português (Brasil)\u003C\u002Fa>\n\u003C\u002Fp>\n\n\u003Ch1 align=\"center\">🦞ClawTeam-OpenClaw\u003C\u002Fh1>\n\n\u003Cp align=\"center\">\n  \u003Cstrong>Multi-agent swarm coordination for CLI coding agents — \u003Ca href=\"https:\u002F\u002Fopenclaw.ai\">OpenClaw\u003C\u002Fa> as default\u003C\u002Fstrong>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n  \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHKUDS\u002FClawTeam\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fupstream-HKUDS%2FClawTeam-purple?style=for-the-badge\" alt=\"Upstream\">\u003C\u002Fa>\n  \u003Ca href=\"#-quick-start\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FQuick_Start-3_min-blue?style=for-the-badge\" alt=\"Quick Start\">\u003C\u002Fa>\n  \u003Ca href=\"LICENSE\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLicense-MIT-yellow?style=for-the-badge\" alt=\"License\">\u003C\u002Fa>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fpython-≥3.10-blue?logo=python&logoColor=white\" alt=\"Python\">\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fagents-OpenClaw_%7C_Claude_Code_%7C_Codex_%7C_Hermes_%7C_nanobot-blueviolet\" alt=\"Agents\">\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Ftransport-File_%7C_ZeroMQ_P2P-orange\" alt=\"Transport\">\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fversion-0.3.0-teal\" alt=\"Version\">\n\u003C\u002Fp>\n\n> **Fork of [HKUDS\u002FClawTeam](https:\u002F\u002Fgithub.com\u002FHKUDS\u002FClawTeam)** with deep OpenClaw integration: default `openclaw` agent, per-agent session isolation, exec approval auto-config, and production-hardened spawn backends. All upstream fixes are synced.\n\nYou set the goal. The agent swarm handles the rest — spawning workers, splitting tasks, coordinating, and merging results.\n\nWorks with [OpenClaw](https:\u002F\u002Fopenclaw.ai) (default), [Claude Code](https:\u002F\u002Fclaude.ai\u002Fclaude-code), [Codex](https:\u002F\u002Fopenai.com\u002Fcodex), [Hermes Agent](https:\u002F\u002Fgithub.com\u002FNousResearch\u002Fhermes-agent), [nanobot](https:\u002F\u002Fgithub.com\u002FHKUDS\u002Fnanobot), [Cursor](https:\u002F\u002Fcursor.com), and any CLI agent.\n\n## Platform Support\n\n- Linux and macOS keep the original `tmux`-first workflow.\n- Windows 10\u002F11 is supported with an automatic fallback to the `subprocess` backend.\n- Task locking, process liveness checks, and signal registration are routed through a shared compatibility layer so unsupported Unix-only behavior degrades safely on Windows.\n- `board attach` still requires `tmux`, so on Windows prefer `clawteam board serve` for live monitoring.\n- If you want the original tmux workflow on Windows, run ClawTeam inside WSL.\n\n---\n\n## Why ClawTeam?\n\nCurrent AI agents are powerful but work in **isolation**. ClawTeam lets agents self-organize into teams — splitting work, communicating, and converging on results without human micromanagement.\n\n| | ClawTeam | Other multi-agent frameworks |\n|---|---------|----------------------------|\n| **Who uses it** | The AI agents themselves | Humans writing orchestration code |\n| **Setup** | `pip install` + one prompt | Docker, cloud APIs, YAML configs |\n| **Infrastructure** | Filesystem + tmux | Redis, message queues, databases |\n| **Agent support** | Any CLI agent | Framework-specific only |\n| **Isolation** | Git worktrees (real branches) | Containers or virtual envs |\n\n---\n\n## How It Works\n\n\u003Ctable>\n\u003Ctr>\n\u003Ctd width=\"33%\">\n\n### Agents Spawn Agents\nThe leader calls `clawteam spawn` to create workers. Each gets its own **git worktree**, **spawn backend session**, and **identity**.\n\n```bash\nclawteam spawn --team my-team \\\n  --agent-name worker1 \\\n  --task \"Implement auth module\"\n```\n\n\u003C\u002Ftd>\n\u003Ctd width=\"33%\">\n\n### Agents Talk to Agents\nWorkers check inboxes, update tasks, and report results — all through CLI commands **auto-injected** into their prompt.\n\n```bash\nclawteam task list my-team --owner me\nclawteam inbox send my-team leader \\\n  \"Auth done. All tests passing.\"\n```\n\n\u003C\u002Ftd>\n\u003Ctd width=\"33%\">\n\n### You Just Watch\nMonitor the swarm from a tiled tmux view or Web UI. The leader handles coordination.\n\n```bash\nclawteam board serve --port 8080\n# Or, on Linux\u002FmacOS\u002FWSL with tmux:\nclawteam board attach my-team\n```\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\n---\n\n## Quick Start\n\n### Option 1: Let the Agent Drive (Recommended)\n\nInstall ClawTeam, then prompt your agent:\n\n```\n\"Build a web app. Use clawteam to split the work across multiple agents.\"\n```\n\nThe agent auto-creates a team, spawns workers, assigns tasks, and coordinates — all via `clawteam` CLI.\n\n### Option 2: Drive It Manually\n\n```bash\n# Create a team\nclawteam team spawn-team my-team -d \"Build the auth module\" -n leader\n\n# Spawn workers — each gets a git worktree plus its own backend session\nclawteam spawn --team my-team --agent-name alice --task \"Implement OAuth2 flow\"\nclawteam spawn --team my-team --agent-name bob   --task \"Write unit tests for auth\"\n\n# Watch them work\nclawteam board serve --port 8080\nclawteam board attach my-team   # Linux\u002FmacOS\u002FWSL with tmux\n```\n\n### Supported Agents\n\n| Agent | Spawn Command | Status |\n|-------|--------------|--------|\n| [OpenClaw](https:\u002F\u002Fopenclaw.ai) | `clawteam spawn --team ...` | **Default** |\n| [Claude Code](https:\u002F\u002Fclaude.ai\u002Fclaude-code) | `clawteam spawn claude --team ...` | Full support |\n| [Codex](https:\u002F\u002Fopenai.com\u002Fcodex) | `clawteam spawn codex --team ...` | Full support |\n| [nanobot](https:\u002F\u002Fgithub.com\u002FHKUDS\u002Fnanobot) | `clawteam spawn nanobot --team ...` | Full support |\n| [Hermes Agent](https:\u002F\u002Fgithub.com\u002FNousResearch\u002Fhermes-agent) | `clawteam spawn hermes --team ...` | Full support (tmux + subprocess) |\n| [Cursor](https:\u002F\u002Fcursor.com) | `clawteam spawn subprocess cursor --team ...` | Experimental |\n| Custom scripts | `clawteam spawn subprocess python --team ...` | Full support |\n\n---\n\n## Install\n\n### Step 1: Prerequisites\n\nClawTeam requires **Python 3.10+** and at least one CLI coding agent (OpenClaw, Claude Code, Codex, etc.). On Linux\u002FmacOS, the full visual workflow also requires **tmux**. On Windows, `tmux` is optional because ClawTeam defaults to the `subprocess` backend.\n\n**Check what you already have:**\n\n```bash\npython --version    # Need 3.10+\ntmux -V             # Linux\u002FmacOS\u002FWSL only\nopenclaw --version  # Or: claude --version \u002F codex --version\n```\n\n**Install missing prerequisites:**\n\n| Tool | Windows | macOS | Ubuntu\u002FDebian |\n|------|---------|-------|---------------|\n| Python 3.10+ | Install from [python.org](https:\u002F\u002Fwww.python.org\u002Fdownloads\u002Fwindows\u002F) | `brew install python@3.12` | `sudo apt update && sudo apt install python3 python3-pip` |\n| tmux | Optional | `brew install tmux` | `sudo apt install tmux` |\n| OpenClaw | `pip install openclaw` | `pip install openclaw` | `pip install openclaw` |\n\n> If using Claude Code or Codex instead of OpenClaw, install those per their own docs. OpenClaw is the default but not strictly required.\n\nOn Windows, after installation you can verify the backend choice with:\n\n```powershell\nclawteam config get default_backend\n```\n\n### Windows Native Setup\n\nUse this path for PowerShell or Windows Terminal:\n\n```powershell\npy -3 -m pip install -e .\nclawteam config get default_backend   # should print subprocess\nclawteam spawn --team demo --agent-name worker1 --task \"Do work\"\nclawteam board serve --port 8080\n```\n\nIf you want the full tmux experience, install and run ClawTeam inside WSL instead.\n\n### Step 2: Install ClawTeam\n\n> **⚠️ Do NOT run `pip install clawteam` or `npm install -g clawteam` directly:**\n> - `pip install clawteam` installs the upstream PyPI version, which defaults to `claude` and lacks OpenClaw adaptations.\n> - `npm install -g clawteam` installs an unrelated name-squatting package (by `a9logic`). If `clawteam --version` shows \"Coming Soon\", you have the wrong one — run `npm uninstall -g clawteam`.\n>\n> **Use the three commands below — the `pip install -e .` step is required. It installs from the local repo, not from PyPI.**\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002Fwin4r\u002FClawTeam-OpenClaw.git\ncd ClawTeam-OpenClaw\npip install -e .    # ← Required! Installs from local repo, NOT the same as pip install clawteam\n```\n\nOptional — P2P transport (ZeroMQ):\n\n```bash\npython -m pip install -e \".[p2p]\"\n```\n\n### Step 3: Ensure `clawteam` is on PATH\n\nSpawned agents run in fresh shells that may not have pip's bin directory in PATH. A symlink in `~\u002Fbin` ensures `clawteam` is always reachable:\n\n```bash\nmkdir -p ~\u002Fbin\nln -sf \"$(which clawteam)\" ~\u002Fbin\u002Fclawteam\n```\n\nIf `which clawteam` returns nothing, find the binary manually:\n\n```bash\n# Common locations:\n# ~\u002F.local\u002Fbin\u002Fclawteam\n# \u002Fopt\u002Fhomebrew\u002Fbin\u002Fclawteam\n# \u002Fusr\u002Flocal\u002Fbin\u002Fclawteam\n# \u002FLibrary\u002FFrameworks\u002FPython.framework\u002FVersions\u002F3.*\u002Fbin\u002Fclawteam\nfind \u002F -name clawteam -type f 2>\u002Fdev\u002Fnull | head -5\n```\n\nThen ensure `~\u002Fbin` is in your PATH — add this to `~\u002F.zshrc` or `~\u002F.bashrc` if it isn't:\n\n```bash\nexport PATH=\"$HOME\u002Fbin:$PATH\"\n```\n\nOn native Windows, you usually do not need the `~\u002Fbin` symlink step. Instead, make sure the Python `Scripts` directory containing `clawteam.exe` is on `PATH`, or activate the virtual environment where you installed ClawTeam before spawning agents.\n\n### Step 4: Install the OpenClaw skill (OpenClaw users only)\n\nThe skill file teaches OpenClaw agents how to use ClawTeam through natural language. Skip this step if you're not using OpenClaw.\n\n```bash\nmkdir -p ~\u002F.openclaw\u002Fworkspace\u002Fskills\u002Fclawteam\ncp skills\u002Fopenclaw\u002FSKILL.md ~\u002F.openclaw\u002Fworkspace\u002Fskills\u002Fclawteam\u002FSKILL.md\n```\n\n### Step 5: Configure exec approvals (OpenClaw users only)\n\nSpawned OpenClaw agents need permission to run `clawteam` commands. Without this, agents will block on interactive permission prompts.\n\n```bash\n# Ensure security mode is \"allowlist\" (not \"full\")\npython3 -c \"\nimport json, pathlib\np = pathlib.Path.home() \u002F '.openclaw' \u002F 'exec-approvals.json'\nif p.exists():\n    d = json.loads(p.read_text())\n    d.setdefault('defaults', {})['security'] = 'allowlist'\n    p.write_text(json.dumps(d, indent=2))\n    print('exec-approvals.json updated: security = allowlist')\nelse:\n    print('exec-approvals.json not found — run openclaw once first, then re-run this step')\n\"\n\n# Add clawteam to the allowlist (use the absolute path — OpenClaw 4.2+ requires it)\nopenclaw approvals allowlist add --agent \"*\" \"$(which clawteam)\"\n```\n\n> If `openclaw approvals` fails, the OpenClaw gateway may not be running. Start it first, then retry.\n\n### Step 5b: Install the Hermes skill (Hermes Agent users only)\n\nThe skill file teaches Hermes Agent how to use ClawTeam through natural language -- including when to route to clawteam (vs `delegate_task`), correct spawn flags, and timing expectations. Skip this step if you're not using Hermes.\n\n```bash\nmkdir -p ~\u002F.hermes\u002Fskills\u002Fopenclaw-imports\u002Fclawteam\ncp skills\u002Fhermes\u002FSKILL.md ~\u002F.hermes\u002Fskills\u002Fopenclaw-imports\u002Fclawteam\u002FSKILL.md\n```\n\n> Verify with `hermes skills list | grep clawteam`. The skill should show up under `openclaw-imports` (Hermes auto-routes skills from that directory).\n\n**Key things the skill teaches Hermes:**\n\n- Route multi-agent\u002Fswarm\u002Fteam queries to clawteam (not `delegate_task`)\n- Use `--team-name` (not `--team`), `-g`\u002F`--goal`, `--force` on `launch`\n- Always pass `--command hermes` on `launch` -- templates default to `openclaw`\n- On `spawn`, pass `hermes` as a trailing positional arg (not `--command hermes`)\n- Wait `sleep 60` after launch for worker boot, then poll the board every 30s\n- Never peek inboxes within the first 60s (they'll be empty)\n- Read inboxes and produce a consolidated report before `clawteam team cleanup`\n\nSpawned Hermes workers automatically inherit MCP servers configured in `~\u002F.hermes\u002Fconfig.yaml`, so any knowledge brain or tool setup is available to every worker.\n\n### Step 6: Verify\n\n```bash\nclawteam --version          # Should print version\nclawteam config health      # Should show all green\n```\n\nIf using OpenClaw, also verify the skill is loaded:\n\n```bash\nopenclaw skills list | grep clawteam\n```\n\n### Automated installer\n\nSteps 2–6 above are also available as a single script:\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002Fwin4r\u002FClawTeam-OpenClaw.git\ncd ClawTeam-OpenClaw\nbash scripts\u002Finstall-openclaw.sh\n```\n\nThis script is intended for Linux, macOS, and WSL shells, not native PowerShell.\n\n### Troubleshooting\n\n| Problem | Cause | Fix |\n|---------|-------|-----|\n| `clawteam: command not found` | pip bin dir not in PATH | Run Step 3 and ensure either `~\u002Fbin` or your Python `Scripts` directory is on PATH |\n| Spawned agents can't find `clawteam` | Agents run in fresh shells without pip PATH | Verify `clawteam` is on PATH in new shells; on Windows check the Python `Scripts` directory or active virtualenv |\n| `openclaw approvals` fails | Gateway not running | Start `openclaw gateway` first, then retry Step 5 |\n| `exec-approvals.json not found` | OpenClaw never ran | Run `openclaw` once to generate config, then retry Step 5 |\n| Agents block on permission prompts | Exec approvals security is \"full\" | Run Step 5 to switch to \"allowlist\" |\n| `pip install -e .` fails | Missing build deps | Run `pip install hatchling` first |\n| `clawteam --version` shows \"Coming Soon\" | Installed the npm name-squatting package (`a9logic`, unrelated to this project) | `npm uninstall -g clawteam`, then reinstall per Step 2 |\n\n---\n\n## Use Cases\n\n### 1. Autonomous ML Research — 8 Agents x 8 GPUs\n\nBased on [@karpathy\u002Fautoresearch](https:\u002F\u002Fgithub.com\u002Fkarpathy\u002Fautoresearch). One prompt launches 8 research agents across H100s that design 2000+ experiments autonomously.\n\n```\nHuman: \"Use 8 GPUs to optimize train.py. Read program.md for instructions.\"\n\nLeader agent:\n├── Spawns 8 agents, each assigned a research direction (depth, width, LR, batch size...)\n├── Each agent gets its own git worktree for isolated experiments\n├── Every 30 min: checks results, cross-pollinates best configs to new agents\n├── Reassigns GPUs as agents finish — fresh agents start from best known config\n└── Result: val_bpb 1.044 → 0.977 (6.4% improvement) across 2430 experiments in ~30 GPU-hours\n```\n\nFull results: [novix-science\u002Fautoresearch](https:\u002F\u002Fgithub.com\u002Fnovix-science\u002Fautoresearch)\n\n### 2. Agentic Software Engineering\n\n```\nHuman: \"Build a full-stack todo app with auth, database, and React frontend.\"\n\nLeader agent:\n├── Creates tasks with dependency chains (API schema → auth + DB → frontend → tests)\n├── Spawns 5 agents (architect, 2 backend, frontend, tester) in separate worktrees\n├── Dependencies auto-resolve: architect completes → backend unblocks → tester unblocks\n├── Agents coordinate via inbox: \"Here's the OpenAPI spec\", \"Auth endpoints ready\"\n└── Leader merges all worktrees into main when complete\n```\n\n### 3. AI Hedge Fund — Template Launch\n\nA TOML template spawns a complete 7-agent investment team with one command:\n\n```bash\nclawteam launch hedge-fund --team fund1 --goal \"Analyze AAPL, MSFT, NVDA for Q2 2026\"\n```\n\nSeven agents total: 5 analysts (value, growth, technical, fundamentals, sentiment) work in parallel, a risk manager synthesizes all signals, and a portfolio manager makes the final decision.\n\nTemplates are TOML files — **create your own** for any domain.\n\n---\n\n## Features\n\n\u003Ctable>\n\u003Ctr>\n\u003Ctd width=\"50%\">\n\n### Agent Self-Organization\n- Leader spawns and manages workers\n- Auto-injected coordination prompt — zero manual setup\n- Workers self-report status and idle state\n- Any CLI agent can participate\n\n### Workspace Isolation\n- Each agent gets its own **git worktree**\n- No merge conflicts between parallel agents\n- Checkpoint, merge, and cleanup commands\n- Branch naming: `clawteam\u002F{team}\u002F{agent}`\n\n### Task Tracking with Dependencies\n- Shared kanban: `pending` → `in_progress` → `completed` \u002F `blocked`\n- `--blocked-by` chains with auto-unblock on completion\n- `task wait` blocks until all tasks complete\n\n\u003C\u002Ftd>\n\u003Ctd width=\"50%\">\n\n### Inter-Agent Messaging\n- Point-to-point inboxes (send, receive, peek)\n- Broadcast to all team members\n- File-based (default) or ZeroMQ P2P transport\n\n### Monitoring & Dashboards\n- `board show` — terminal kanban\n- `board live` — auto-refreshing dashboard\n- `board attach` — tiled tmux view of all agents (Linux\u002FmacOS\u002FWSL)\n- `board serve` — Web UI with real-time updates\n\n### Team Templates\n- TOML files define team archetypes (roles, tasks, prompts)\n- One command: `clawteam launch \u003Ctemplate>`\n- Variable substitution: `{goal}`, `{team_name}`, `{agent_name}`\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\n### v0.3.0 — Production Intelligence *(New)*\n- **Hermes Agent Support** — native spawn target across NativeCliAdapter, tmux, and subprocess backends. Auto-inserts `chat` subcommand and passes `--source tool` (session hygiene requires the upstream Hermes patch described in `skills\u002Fhermes\u002FSKILL.md` — ClawTeam passes the flag correctly; Hermes ≤ 0.8.0 ignores it).\n- **Cost Dashboard** — real-time token\u002Fcost by agent, model, and task (`clawteam board cost`). No competitor has this.\n- **Circuit Breaker** — healthy → degraded → open tri-state with half-open probing\n- **Retry with Backoff** — `spawn_with_retry()` for resilient agent spawning\n- **Idempotency Keys** — deduplication for `create()` and `send()`\n- **Intent-Based Prompts** — military C2 Auftragstaktik: agents get `intent` + `end_state` + `constraints`\n- **Boids Emergence Rules** — Reynolds 1986 flocking rules adapted for LLM agents\n- **Metacognitive Self-Assessment** — agents tag their own confidence levels\n- **Per-Agent Model Resolution** — 7-level priority chain, mix Claude\u002FGPT\u002FQwen in one team\n- **Runtime Live Injection** — `runtime inject\u002Fstate\u002Fwatch` for messaging running agents\n\n**Also:** plan approval workflows, graceful lifecycle management, `--json` output on all commands, cross-machine support (NFS\u002FSSHFS or P2P), multi-user namespacing, spawn validation with auto-rollback, `fcntl` file locking for concurrent safety.\n\n---\n\n## OpenClaw Integration\n\nThis fork makes [OpenClaw](https:\u002F\u002Fopenclaw.ai) the **default agent**. Without ClawTeam, each OpenClaw agent works in isolation. ClawTeam transforms it into a multi-agent platform.\n\n| Capability | OpenClaw Alone | OpenClaw + ClawTeam |\n|-----------|---------------|-------------------|\n| **Task assignment** | Manual per-agent messaging | Leader autonomously splits, assigns, monitors |\n| **Parallel development** | Shared working directory | Isolated git worktrees per agent |\n| **Dependencies** | Manual polling | `--blocked-by` with auto-unblock |\n| **Communication** | Only through AGI relay | Direct point-to-point inbox + broadcast |\n| **Observability** | Read logs | Kanban board + tiled tmux view |\n\nOnce the skill is installed, talk to your OpenClaw bot in any channel:\n\n| What you say | What happens |\n|-------------|-------------|\n| \"Create a 5-agent team to build a web app\" | Creates team, tasks, and spawns 5 agents with the configured backend |\n| \"Launch a hedge-fund analysis team\" | `clawteam launch hedge-fund` with 7 agents |\n| \"Check the status of my agent team\" | `clawteam board show` with kanban output |\n\n```\n  You (Telegram\u002FDiscord\u002FTUI)\n         │\n         ▼\n  ┌──────────────────┐\n  │  OpenClaw Gateway │  ← activates clawteam skill\n  └────────┬─────────┘\n           │\n           ▼\n  ┌──────────────────┐     clawteam spawn     ┌─────────────────┐\n  │  Leader Agent    │ ─────────────────────► │  openclaw tui   │\n  │  (openclaw)      │ ──┐                    │  (tmux window)  │\n  │                  │   │                    │  git worktree   │\n  │  Manages swarm   │   ├──────────────────► ├─────────────────┤\n  │  via clawteam    │   │                    │  openclaw tui   │\n  │  CLI             │   ├──────────────────► ├─────────────────┤\n  └──────────────────┘   │                    │  openclaw tui   │\n                         └──────────────────► └─────────────────┘\n                                               All coordinate via\n                                               ~\u002F.clawteam\u002F (tasks, inboxes)\n```\n\n---\n\n## Hermes Agent Integration\n\nClawTeam ships first-class support for [Hermes Agent](https:\u002F\u002Fgithub.com\u002FNousResearch\u002Fhermes-agent) — Nous Research's self-improving CLI agent. Hermes workers spawn via the same adapter path as OpenClaw (tmux or subprocess), but use Hermes-native command flags (`hermes chat --yolo --source tool -q \"\u003Ctask>\"`).\n\nHermes workers automatically inherit any MCP servers configured in `~\u002F.hermes\u002Fconfig.yaml`, so whatever tools you've wired into Hermes are available to every spawned worker.\n\n| Capability | Hermes Alone | Hermes + ClawTeam |\n|-----------|--------------|-------------------|\n| **Parallelism** | Single session | Spawn N workers in tmux windows |\n| **Coordination** | Manual | Kanban + inboxes + task dependencies |\n| **Isolation** | Shared working dir | Git worktrees per agent |\n| **Session hygiene** | Mixes with user sessions | `--source tool` tag passed to Hermes (requires upstream fix — see SKILL.md `Known upstream issues`) |\n\n**Using Hermes with ClawTeam:**\n\nAll built-in templates (`hedge-fund`, `research-paper`, `code-review`, `strategy-room`) default to spawning OpenClaw workers. Hermes users pass `--command hermes` to override:\n\n```bash\nclawteam launch hedge-fund --team-name \u003Cname> --goal \"...\" --command hermes --force\n```\n\nOr spawn manually, passing `hermes` as the trailing positional argument:\n\n```bash\nclawteam spawn --team \u003Cteam> --agent-name \u003Cname> --task \"...\" --no-workspace hermes\n```\n\nNote: the built-in templates were designed around OpenClaw's `clawteam inbox send` coordination pattern. Hermes workers sometimes complete their analysis without executing the inbox-send command. If `clawteam inbox peek` returns empty while the kanban shows `COMPLETED`, capture tmux scrollback directly:\n\n```bash\ntmux capture-pane -t clawteam-\u003Cteam>:\u003Cwindow-index> -p -S -500\n```\n\n**Installation:** see Step 5b in the Install section.\n\n---\n\n## Architecture\n\n```\n  Human: \"Optimize this LLM\"\n         │\n         ▼\n  ┌──────────────┐     clawteam spawn     ┌──────────────┐\n  │  Leader      │ ──────────────────────► │  Worker      │\n  │  (any agent) │ ──────┐                │  git worktree │\n  │              │       ├──────────────► │  tmux window  │\n  │  spawn       │       │                ├──────────────┤\n  │  task create │       ├──────────────► │  Worker      │\n  │  inbox send  │       │                │  git worktree │\n  │  board show  │       └──────────────► │  tmux window  │\n  └──────────────┘                        └──────────────┘\n                                                 │\n                                                 ▼\n                                      ┌─────────────────────┐\n                                      │    ~\u002F.clawteam\u002F     │\n                                      │ ├── teams\u002F   (who) │\n                                      │ ├── tasks\u002F   (what)│\n                                      │ ├── inboxes\u002F (talk)│\n                                      │ └── workspaces\u002F    │\n                                      └─────────────────────┘\n```\n\nAll state lives in `~\u002F.clawteam\u002F` as JSON files. No database, no server. Atomic writes with cross-platform file locking ensure crash safety.\n\n| Setting | Env Var | Default |\n|---------|---------|---------|\n| Data directory | `CLAWTEAM_DATA_DIR` | `~\u002F.clawteam` |\n| Transport | `CLAWTEAM_TRANSPORT` | `file` |\n| Workspace mode | `CLAWTEAM_WORKSPACE` | `auto` |\n| Spawn backend | `CLAWTEAM_DEFAULT_BACKEND` | `tmux` on Linux\u002FmacOS, `subprocess` on Windows |\n\n---\n\n## Command Reference\n\n\u003Cdetails open>\n\u003Csummary>\u003Cstrong>Core Commands\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n```bash\n# Team lifecycle\nclawteam team spawn-team \u003Cteam> -d \"description\" -n \u003Cleader>\nclawteam team discover                    # List all teams\nclawteam team status \u003Cteam>               # Show members\nclawteam team cleanup \u003Cteam> --force      # Delete team\n\n# Spawn agents (note: `spawn` uses --team; `launch` uses --team-name)\nclawteam spawn --team \u003Cteam> --agent-name \u003Cname> --task \"do this\"\nclawteam spawn codex --team \u003Cteam> --agent-name \u003Cname> --task \"do this\"\nclawteam spawn --team \u003Cteam> --agent-name \u003Cname> --task \"do this\" hermes\nclawteam spawn subprocess hermes --team \u003Cteam> --agent-name \u003Cname> --task \"do this\"\n\n# Task management\nclawteam task create \u003Cteam> \"subject\" -o \u003Cowner> --blocked-by \u003Cid1>,\u003Cid2>\nclawteam task update \u003Cteam> \u003Cid> --status completed   # auto-unblocks dependents\nclawteam task list \u003Cteam> --status blocked --owner worker1\nclawteam task wait \u003Cteam> --timeout 300\n\n# Messaging\nclawteam inbox send \u003Cteam> \u003Cto> \"message\"\nclawteam inbox broadcast \u003Cteam> \"message\"\nclawteam inbox receive \u003Cteam>             # consume messages\nclawteam inbox peek \u003Cteam>                # read without consuming\n\n# Monitoring\nclawteam board show \u003Cteam>                # terminal kanban\nclawteam board live \u003Cteam> --interval 3   # auto-refresh\nclawteam board attach \u003Cteam>              # tiled tmux view (Linux\u002FmacOS\u002FWSL)\nclawteam board serve --port 8080          # web UI\n```\n\n\u003C\u002Fdetails>\n\n\u003Cdetails>\n\u003Csummary>\u003Cstrong>Workspace, Plan, Lifecycle, Config\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n```bash\n# Workspace (git worktree management)\nclawteam workspace list \u003Cteam>\nclawteam workspace checkpoint \u003Cteam> \u003Cagent>    # auto-commit\nclawteam workspace merge \u003Cteam> \u003Cagent>         # merge back to main\nclawteam workspace cleanup \u003Cteam> \u003Cagent>       # remove worktree\n\n# Plan approval\nclawteam plan submit \u003Cteam> \u003Cagent> \"plan\" --summary \"TL;DR\"\nclawteam plan approve \u003Cteam> \u003Cplan-id> \u003Cagent> --feedback \"LGTM\"\nclawteam plan reject \u003Cteam> \u003Cplan-id> \u003Cagent> --feedback \"Revise X\"\n\n# Lifecycle\nclawteam lifecycle request-shutdown \u003Cteam> \u003Cagent> --reason \"done\"\nclawteam lifecycle approve-shutdown \u003Cteam> \u003Crequest-id> \u003Cagent>\nclawteam lifecycle idle \u003Cteam>\n\n# Templates\nclawteam launch \u003Ctemplate> --team \u003Cname> --goal \"Build X\"\nclawteam template list\n\n# Config\nclawteam config show\nclawteam config set transport p2p\nclawteam config health\n```\n\n\u003C\u002Fdetails>\n\n---\n\n## Per-Agent Model Assignment\n\nAssign different models to different agent roles for better cost\u002Fperformance tradeoffs in multi-agent swarms. Uses a **7-level priority chain**: CLI > agent model > agent tier > template strategy > template model > config default > None.\n\n**Per-agent model in templates:**\n```toml\n[template]\nname = \"my-team\"\ncommand = [\"openclaw\"]\nmodel = \"sonnet-4.6\"              # default for all agents\nmodel_strategy = \"auto\"           # or: leaders→strong, workers→balanced\n\n[template.leader]\nname = \"lead\"\nmodel = \"opus\"                    # override for leader\n\n[[template.agents]]\nname = \"worker\"\nmodel_tier = \"cheap\"              # cost tiers: strong \u002F balanced \u002F cheap\n```\n\n**CLI flags:**\n```bash\nclawteam spawn --model opus                          # single agent\nclawteam launch my-template --model gpt-5.4          # override all agents\nclawteam launch my-template --model-strategy auto     # auto-assign by role\n```\n\n---\n\n\n## Roadmap\n\n| Version | What | Status |\n|---------|------|--------|\n| v0.2 | OpenClaw default agent, workspace overlay, zombie detection, 11-language README | Shipped |\n| v0.3 | Research-backed intelligence, cost dashboard, circuit breaker, per-agent models, runtime injection | **Shipped** |\n| v0.4 | Windows full support, A2A Gateway integration | In Progress |\n| v0.5 | Agent template marketplace — community-contributed TOML templates | Planned |\n| v0.6 | Memory deep integration — per-team\u002Fper-task knowledge sharing | Planned |\n| v1.0 | Production-grade — auth, permissions, audit logs | Exploring |\n\n---\n\n## Contributing\n\nWe welcome contributions! See [CONTRIBUTING.md](CONTRIBUTING.md) for setup, code style, and PR guidelines.\n\nAreas we'd love help with:\n\n- **Agent integrations** — support for more CLI agents\n- **Team templates** — TOML templates for new domains\n- **Transport backends** — Redis, NATS, etc.\n- **Dashboard improvements** — Web UI, Grafana\n- **Documentation** — tutorials and best practices\n\n---\n\n## Acknowledgements\n\n- [@karpathy\u002Fautoresearch](https:\u002F\u002Fgithub.com\u002Fkarpathy\u002Fautoresearch) — autonomous ML research framework\n- [OpenClaw](https:\u002F\u002Fopenclaw.ai) — default agent backend\n- [Hermes Agent](https:\u002F\u002Fgithub.com\u002FNousResearch\u002Fhermes-agent) — Nous Research's self-improving CLI agent\n- [Claude Code](https:\u002F\u002Fclaude.ai\u002Fclaude-code) and [Codex](https:\u002F\u002Fopenai.com\u002Fcodex) — supported AI coding agents\n- [ai-hedge-fund](https:\u002F\u002Fgithub.com\u002Fvirattt\u002Fai-hedge-fund) — hedge fund template inspiration\n- [CLI-Anything](https:\u002F\u002Fgithub.com\u002FHKUDS\u002FCLI-Anything) — sister project\n\n## License\n\nMIT — free to use, modify, and distribute.\n\n---\n\n\u003Cdiv align=\"center\">\n\n**ClawTeam** — *Agent Swarm Intelligence.*\n\n\u003C\u002Fdiv>\n","ClawTeam-OpenClaw 是一个专为多智能体群体协调设计的项目，以 OpenClaw 作为默认智能体。该项目支持多种智能体如 Claude Code、Codex 和 Hermes Agent 等，并通过 File 或 ZeroMQ P2P 方式进行通信。其核心功能包括任务分配、智能体间的协作与结果整合，特别强调了对 OpenClaw 的深度集成，如会话隔离和执行审批自动配置等特性。适用于需要多个编程助手协同工作的场景，例如大规模代码生成或复杂软件开发流程自动化。此外，该工具在 Linux、macOS 以及 Windows 10\u002F11 上均能良好运行，尽管某些高级功能可能需要特定环境如 tmux 或 WSL 支持。","2026-06-11 03:49:39","high_star"]