One-command multi-agent initialization for OpenClaw
Ship an entire AI agent fleet to your chat group in 60 seconds. Default model: zai/glm-5

Installation • Architecture • Agents • Channels • Workflows • Customization • Contributing

OpenClaw Agents is a ready-to-deploy multi-agent configuration kit for OpenClaw. It provisions 9 specialized AI agents as a collaborative team — complete with identities, workspace files, routing rules, and channel bindings — using a single setup command.

• 🤖 9 pre-configured agents with rich emoji identities for instant recognition in chat
• 📝 Agent self-merge — deploys BOOTSTRAP.md for intelligent first-run setup
• 🔗 Two Deployment Modes:
  • Channel Mode: auto-routes agents to Feishu, WhatsApp, Telegram, Discord (supports individual groups)
  • Local Workflow Mode: no channel required, agents communicate directly via agentToAgent tool
• 📋 4 workflow templates — Paper Pipeline, Daily Digest, Brainstorm, Rebuttal
• ⚔️ Adversarial collaboration — Ideator↔Critic and Writer↔Reviewer creative tension
• 🎯 SHARP taste gates — quality checkpoints enforced by the Critic agent
• 🛡️ Safe merge — appends agents to existing config, never overwrites your main agent
• 🎛️ Flexible model — unified model or per-agent customization via --model-map
• 🧩 Extensible — add custom agents via agents.yaml

Paste this into your LLM agent session (Claude Code, Antigravity, Cursor, etc.):

Fetch the OpenClaw Agents installation guide and set up the multi-agent system by following:
curl -s https://raw.githubusercontent.com/shenhao-stu/openclaw-agents/main/docs/installation.md

Or just run it yourself:

git clone https://github.com/shenhao-stu/openclaw-agents.git
cd openclaw-agents
chmod +x setup.sh
./setup.sh

IMPORTANT: Use curl to fetch this file, NOT WebFetch. WebFetch summarizes content and loses critical details like CLI flags and config examples. Always use:

curl -fsSL https://raw.githubusercontent.com/shenhao-stu/openclaw-agents/main/docs/installation.md

If you're an LLM Agent helping set up OpenClaw Agents, follow the full installation guide. Here's the summary:

# Verify OpenClaw CLI is installed
if command -v openclaw &> /dev/null; then
    echo "OpenClaw $(openclaw --version) is installed"
else
    echo "Installing OpenClaw..."
    npm install -g openclaw@latest
    openclaw onboard --install-daemon
fi

Ask the user these questions:

1. Which deployment mode do you want?
   • Mode 1 (Channel Mode): Deploy agents to Feishu, Slack, Telegram, etc.
   • Mode 2 (Local Workflow Mode): Use locally via CLI workflows. Agents talk to each other via the agentToAgent tool. (No channel needed).

(If Mode 1: Channel Mode) 2. Which channel? → --channel feishu|whatsapp|... 3. How to assign groups?

• All in one group: provide one --group-id <ID>
• Separate groups per agent: interactively paste 8 different group IDs in the script.

4. 需要 @mention 才回复吗？ → --require-mention true|false

(For both modes) 5. Which LLM model? (default: zai/glm-5) → --model <MODEL>

• Different models per agent? → --model-map 'coder=ollama/kimi-k2.5:cloud'

git clone https://github.com/shenhao-stu/openclaw-agents.git
cd openclaw-agents
chmod +x setup.sh
./setup.sh --channel <CHANNEL> --group-id <GROUP_ID>

Examples:

• Interactive Setup (Highly Recommended):

  ./setup.sh

  (The script will elegantly ask you your mode, your channel, and let you paste your 8 group IDs one by one if desired).

• Local Workflow Mode:

  ./setup.sh --mode local

• All agents in one Feishu group:

  ./setup.sh --mode channel --channel feishu --group-id oc_xxx

• Scripted per-agent groups:

  ./setup.sh --mode channel --channel feishu --group-id oc_default \
    --group-map 'coder=oc_dev,scout=oc_news'

• Custom models + no @mention:

  ./setup.sh --channel feishu --group-id oc_xxx \
    --model-map 'coder=ollama/kimi-k2.5:cloud' \
    --require-mention false

• Agents only (no channel): ./setup.sh --skip-bindings

• Dry-run preview: ./setup.sh --dry-run --channel feishu --group-id oc_xxx

The script will:

1. ✅ Verify openclaw CLI is installed
2. 🤖 Create 8 sub-agents via openclaw agents add (auto-generates AGENTS.md, SOUL.md, USER.md)
3. 🎨 Set emoji identities via openclaw agents set-identity
4. 📝 Deploy source files + BOOTSTRAP.md for agent self-merge on first run
5. 📋 Append workflow instructions to each agent's AGENTS.md
6. 🔗 Configure openclaw.json with channel bindings
7. ✅ Verify the entire setup

openclaw agents list --bindings    # Should show all 8 agents with channel bindings
openclaw channels status --probe   # Should show channel connected

openclaw gateway

Then mention any agent in your chat group to test. Each agent will respond with its distinct emoji identity.

⚠️ Warning: Do not modify the 8 core agent IDs (planner, ideator, critic, surveyor, coder, writer, reviewer, scout). These are protected and referenced throughout the workflow system.

                         ┌──────────────┐
                         │   👤 User    │
                         └──────┬───────┘
                                │
                    ┌───────────▼───────────┐
                    │  🐾 OpenClaw Main     │
                    │  (Audit · Manage · Arbitrate)
                    └───────────┬───────────┘
                                │
                         ┌──────▼───────┐
                         │  🧠 Planner  │ ◄── Orchestration Hub
                         └──────┬───────┘
                                │
       ┌────────────────────────┼────────────────────────┐
       │                        │                        │
 ┌─────▼─────┐           ┌─────▼─────┐           ┌─────▼─────┐
 │ 💡Ideator │◄── ⚔️ ──►│ 🎯 Critic │           │ 📰 Scout  │
 │  Creative  │           │   Taste    │           │   Intel    │
 └─────┬─────┘           └─────┬─────┘           └───────────┘
       │                       │
 ┌─────▼─────┐           ┌─────▼─────┐
 │ 📚Surveyor│           │ 💻 Coder  │
 │  Research  │           │ Engineer   │
 └─────┬─────┘           └─────┬─────┘
       │                       │
       └───────────┬───────────┘
             ┌─────▼─────┐
             │ ✍️ Writer │
             │  Author    │
             └─────┬─────┘
                   │
             ┌─────▼─────┐
             │ 🔍Reviewer│ ◄── Quality Gate
             │  Reviewer  │
             └───────────┘

The system is built on productive tension between agents:

• 🎯 Critic holds ultimate taste veto — no idea passes Phase 2.5 without SHARP ≥ 18
• 🔍 Reviewer holds ultimate quality veto — paper cannot submit without Reviewer's Accept

Each agent has three core files inside .agents/<agent_id>/:

Docs: Groups · Multi-Agent Routing

OpenClaw uses a three-tier access control model for groups:

{
  "channels": {
    "feishu": {
      "groupPolicy": "open",                    // ← all groups allowed
      "groups": {
        "oc_YOUR_GROUP_ID": {
          "requireMention": true               // ← must @mention to trigger
        }
      }
    }
  }
}

You can choose whether agents require @mention to respond in groups:

# Default: require @mention
./setup.sh --channel feishu --group-id oc_xxx

# Auto-respond without @mention
./setup.sh --require-mention false --channel feishu --group-id oc_xxx

Each agent has unique mentionPatterns:

{
  "agents": {
    "list": [
      {
        "id": "planner",
        "name": "💡 Planner",
        "groupChat": {
          "mentionPatterns": ["@planner", "planner", "@Planner"],
          "historyLimit": 50
        }
      }
    ]
  }
}

How it works: Type @planner 请分解这个任务 in the group, and only the 💡 Planner agent will respond.

Messages that don't match any mention pattern are stored for context but don't trigger a reply — this allows agents to follow the conversation passively.

Each agent gets an isolated session per group:

agent:<agentId>:<channel>:group:<groupId>

Telegram forum topics add :topic:<threadId> for per-topic isolation.

Each sub-agent has its own independent workspace (no Docker sandbox). The setup script deploys source files that the agent merges on first run:

You can restrict tools per group or per sender:

{
  "channels": {
    "telegram": {
      "groups": {
        "-1001234567890": {
          "tools": {
            "deny": ["exec", "write"]                   // block risky tools
          },
          "toolsBySender": {
            "id:123456789": { "alsoAllow": ["exec"] }   // override for trusted user
          }
        }
      }
    }
  }
}

Agents show as <emoji> <name> in chat (configured via identity.name):

The Critic agent enforces quality at four critical checkpoints:

The default model is zai/glm-5. You have three options:

# Uses zai/glm-5 for all 8 sub-agents
./setup.sh --channel feishu --group-id oc_xxx

# All agents use the same custom model
./setup.sh --model ollama/kimi-k2.5:cloud --channel feishu --group-id oc_xxx

# Default is zai/glm-5, but coder and writer get different models
./setup.sh \
  --model zai/glm-5 \
  --model-map 'coder=ollama/kimi-k2.5:cloud,writer=zai/glm-4.7,scout=zai/glm-4.7-flash' \
  --channel feishu --group-id oc_xxx

--model-map takes priority over --model for specified agents. Agents not in the map use the --model default.

1. Add the agent definition to agents.yaml:

agents:
  # ... existing agents ...
  - id: "math-prover"
    name: "🔢 Math Prover"
    emoji: "🔢"
    role: "Theorem proving, convergence analysis"
    model: "zai/glm-5"
    protected: false
    workspace: ".agents/math-prover"

2. Re-run ./setup.sh or add manually:

openclaw agents add math-prover --model zai/glm-5 --workspace .agents/math-prover
openclaw agents set-identity --agent math-prover --name "🔢 Math Prover"

openclaw-agents/
├── setup.sh                          # 🚀 One-command setup script
├── agents.yaml                       # 📋 Agent manifest (source of truth)
├── soul.md                           # 🐾 Main Agent definition
├── README.md                         # 📖 This file
├── LICENSE                           # MIT License
├── CONTRIBUTING.md                   # Contribution guidelines
├── CHANGELOG.md                      # Version history
├── docs/
│   └── installation.md               # 📖 Full installation guide
├── examples/
│   ├── openclaw.feishu.json          # Feishu config example
│   ├── openclaw.whatsapp.json        # WhatsApp config example
│   └── openclaw.telegram.json        # Telegram config example
└── .agents/
    ├── planner/                      # 🧠 soul.md + agent.md + user.md
    ├── ideator/                      # 💡 soul.md + agent.md + user.md
    ├── critic/                       # 🎯 soul.md + agent.md + user.md
    ├── surveyor/                     # 📚 soul.md + agent.md + user.md
    ├── coder/                        # 💻 soul.md + agent.md + user.md
    ├── writer/                       # ✍️ soul.md + agent.md + user.md
    ├── reviewer/                     # 🔍 soul.md + agent.md + user.md
    ├── scout/                        # 📰 soul.md + agent.md + user.md
    └── workflows/
        ├── paper-pipeline.md         # 📋 End-to-end paper workflow
        ├── daily-digest.md           # 📰 Daily paper digest
        ├── brainstorm.md             # 💡 Idea brainstorming
        └── rebuttal.md               # 🔄 Rebuttal preparation

🛡️ Safe Merge: The setup script appends sub-agents to your existing openclaw.json. Main agent is implicit (uses agents.defaults). No sandbox — each agent has its own workspace. A backup is created automatically.

openclaw agents list --bindings       # List all agents and bindings
openclaw agents add <id>              # Add a new agent
openclaw agents set-identity          # Set agent display name
openclaw channels status --probe      # Check channel connectivity
openclaw gateway                      # Start the gateway
openclaw gateway restart              # Restart after config changes

We welcome contributions! See CONTRIBUTING.md for guidelines.

• 🐛 Bug Reports — Open a GitHub Issue
• 💡 New Agents — Submit a PR with agent files + agents.yaml update
• 📋 Workflows — Share your research process templates
• 📖 Docs — Improve guides and examples

MIT — Use freely, modify openly, share generously.

Built with ❤️ for the AI research community
_{Powered by OpenClaw 🦞}