OpenClaw Task Router 6 Specialist Models Auto Fallback

Production-grade multi-agent orchestration for OpenClaw
Route tasks to the best AI model. Get real-time status. Never wait at a blank terminal.

Quick Start · Agents · How It Works · Customize · Gist Mirror

A single prompt + config that turns your OpenClaw instance into an AI command center. Instead of one model doing everything poorly, 6 specialist agents handle what they're best at:

All models served via OpenCode — one API, one subscription, no per-provider keys needed.

• 🧠 Smart Task Classification — automatically routes to the right specialist
• ⚡ Async Delegation — spawns agents and returns instantly; no waiting
• 📊 2-Minute Heartbeat — live status trees showing what's running, blocked, or done
• 🔄 Automatic Fallback — when a model hits its rate limit, switches to the next best
• 🛡️ Hallucination Prevention — hard NEVER/ALWAYS constraints from production experience
• 🔁 Loop Detection — catches stuck agents and intervenes
• 💰 Cost Optimization — cheap models for simple tasks, powerful reasoning for complex ones

npm install -g openclaw
openclaw onboard

See the OpenClaw docs for full setup.

# Clone this repo or download directly
mkdir -p ~/.openclaw/workspace
curl -o ~/.openclaw/workspace/AGENTS.md https://raw.githubusercontent.com/FaridLU/openclaw-task-router/main/AGENTS.md

Where does OpenClaw look for AGENTS.md? OpenClaw loads AGENTS.md from the workspace directory at session start. Default: ~/.openclaw/workspace/. Configure it in openclaw.json:

{ "agents": { "defaults": { "workspace": "/path/to/your/workspace" } } }

Add the openclaw.json from this repo to your OpenClaw config:

cp openclaw.json ~/.openclaw/openclaw.json
# Then edit to add your OpenCode API key and any personal preferences

Or manually add the 6 specialist agents to your existing config — see openclaw.json for the full template.

Just talk to OpenClaw normally:

Plan a microservices architecture for our e-commerce app

Debug the 500 error in production

Review PR #42

The router announces the chosen agent and model, then returns instantly. Every 2 minutes you'll see:

Router
├── plan (Microservices architecture) → GLM-5.1 (thinking: high) [RUNNING]
└── fast (Search session handling) → DeepSeek V4 Flash (thinking: off) [FINISHED: Found in auth/views.py]

All models via OpenCode (opencode-go/) provider — opencode.ai

User → Task Router → Classify → Spawn Specialist Agent → Return to User
 ↓
 ┌─────────────────────────────────────┐
 │ plan → GLM-5.1 (high thinking) │
 │ build → Kimi K2.6 (high thinking) │
 │ debug → DeepSeek V4 Pro (medium) │
 │ review → Kimi K2.6 (high thinking) │
 │ visual → MiMo V2 Omni (no thinking) │
 │ fast → DeepSeek V4 Flash (no thinking)│
 └─────────────────────────────────────┘
 ↓
 Agent finishes → Result appears in session

If the primary model is unavailable (rate limit, outage, etc.):

plan: GLM-5.1 → Kimi K2.6
build: Kimi K2.6 → DeepSeek V4 Pro → Qwen 3.6 Plus
debug: DeepSeek V4 Pro → Kimi K2.6 → DeepSeek V4 Flash
review: Kimi K2.6 → DeepSeek V4 Pro
visual: MiMo V2 Omni → Qwen 3.6 Plus
fast: DeepSeek V4 Flash → Qwen 3.5 Plus

You never see the switch. The router handles it and reports fallbacks in the next heartbeat.

Replace any model in openclaw.json with your preferred provider:

{
 "id": "plan",
 "model": {
 "primary": "anthropic/claude-sonnet-4-20250514",
 "fallbacks": ["opencode-go/kimi-k2.6"]
 }
}

Works with any OpenRouter, OpenAI, Anthropic, or local model provider OpenClaw supports.

Define new specialists in openclaw.json and add routing rules to AGENTS.md:

{
 "id": "security",
 "description": "Security audits and vulnerability scanning",
 "model": { "primary": "opencode-go/deepseek-v4-pro" }
}

Change high → medium → off per agent to trade reasoning depth for speed/cost:

{ "id": "fast", "thinking": "off" } // cheapest, fastest
{ "id": "debug", "thinking": "medium" } // balanced
{ "id": "plan", "thinking": "high" } // deepest reasoning

openclaw-task-router/
├── AGENTS.md ← The router prompt (copy to ~/.openclaw/workspace/)
├── openclaw.json ← Full agent configuration (copy to ~/.openclaw/)
├── README.md ← This file
├── LICENSE ← MIT
├── CHANGELOG.md ← Version history
└── examples/
 └── demo-session.md ← Example router conversation

1. Fork this repo
2. Create your feature branch (git checkout -b feature/my-improvement)
3. Commit your changes
4. Push to the branch (git push origin feature/my-improvement)
5. Open a Pull Request

This is a living project. Every time the router makes a mistake, update the Anti-Patterns section in AGENTS.md to encode the lesson.

• Farid (@FaridLU) — OpenClaw power user who wanted the ultimate AI orchestration experience
• Claw 🔀 — The OpenClaw Task Router agent that eats its own dogfood

Powered by OpenCode models — one API, six specialists, zero per-provider keys.

MIT — use it, fork it, share it.