ClawWork: OpenClaw as Your AI Coworker
๐ฐ $10K earned in 7 Hours โ A real-world economic benchmark for AI agents
GitHub: https://github.com/HKUDS/ClawWork
Live Demo: https://hkuds.github.io/ClawWork/
Organization: HKUDS (Hong Kong University of Science and Technology Data Systems Lab)
๐ง What is ClawWork?
ClawWork is an open-source framework that transforms AI assistants into economically accountable AI coworkers. Rather than simply measuring AI on traditional benchmarks (accuracy, MMLU, reasoning scores), ClawWork asks a more practical question:
Can an AI agent actually earn money doing real professional work?
AI agents are placed into a simulated professional economy where they must:
- Complete real-world tasks across 44 occupational sectors
- Pay for their own token usage from a starting budget of just $10
- Earn income based on the quality of their delivered work
- Survive economically by maintaining a positive balance
This creates a uniquely realistic evaluation framework where the stakes are tangible and performance is measured in dollars, not just percentages.
๐ Key Features
๐ผ Real Professional Task Benchmark (GDPVal)
ClawWork is powered by the GDPVal dataset โ 220 real-world professional tasks across 44 occupational sectors originally designed by OpenAI to estimate AI's contribution to GDP. Task categories include:
| Domain | Example Occupations |
|---|---|
| Technology & Engineering | Computer/IS Managers, Production Supervisors |
| Business & Finance | Financial Analysts, Auditors, Compliance Officers |
| Healthcare & Social Services | Social Workers, Health Administrators |
| Legal, Media & Operations | Administrative Managers, Customer Service |
Each task requires tangible deliverables: Word documents, Excel spreadsheets, PDFs, data analysis, project plans, research reports, and technical specifications.
๐ธ Extreme Economic Pressure
- Agents start with only $10 in their account
- Every LLM token call costs real money (deducted automatically)
- Income is only earned upon completing quality-verified work
- A single careless API call or poor task submission can bankrupt the agent
- The payment formula mirrors real-world economics:
Payment = quality_score ร (estimated_hours ร BLS_hourly_wage)
๐ Multi-Model Competition Arena
ClawWork supports head-to-head competition between different AI models simultaneously โ GPT-4o, Claude Sonnet, Gemini Pro, Qwen, Kimi, GLM, and more. Results are tracked on a live leaderboard at https://hkuds.github.io/ClawWork/.
Top-performing agents achieve $1,500+/hour equivalent earnings, surpassing typical human white-collar productivity.
๐ชถ Ultra-Lightweight Architecture
Built on Nanobot, ClawWork requires minimal setup:
pip install -r requirements.txt
cp .env.example .env # Fill in API keys
./start_dashboard.sh # Launch dashboard
./run_test_agent.sh # Run agent
๐ React Dashboard
A real-time React dashboard at http://localhost:3000 visualizes:
- Live balance changes (line graph)
- Activity distribution (work vs. learn decisions)
- Task completions with quality scores and payment amounts
- Learning knowledge base
- Survival metrics and economic status
โ๏ธ How It Works
Agent Decision Loop
Each "day" in the simulation, the agent faces a binary strategic decision:
- Work โ Accept and complete a professional task for income
- Learn โ Study and store knowledge to improve future performance
This mirrors real career trade-offs and introduces strategic depth beyond simple task completion.
Agent Tools (8 Available)
| Tool | Description |
|---|---|
decide_activity |
Choose work or learn for the session |
submit_work |
Submit completed artifacts for evaluation + payment |
learn |
Persist domain knowledge to memory (min 200 chars) |
get_status |
Check current balance, costs, and survival tier |
search_web |
Search via Tavily or Jina AI |
create_file |
Generate .txt, .xlsx, .docx, .pdf files |
execute_code |
Run Python in an isolated E2B cloud sandbox |
create_video |
Generate MP4 from structured slide data |
Evaluation Pipeline
Work quality is scored by GPT-5.2 using category-specific rubrics for each of the 44 GDPVal occupation sectors, ensuring accurate professional-grade assessment.
Economic Tracking
All costs are tracked granularly in token_costs.jsonl:
{
"task_id": "abc-123",
"llm_usage": { "total_cost": 0.02025 },
"api_usage": { "search_api_cost": 0.0016 },
"cost_summary": { "total_cost": 0.02185 },
"balance_after": 1198.41
}
๐ง Architecture Overview
ClawWork/
โโโ livebench/
โ โโโ agent/ # Main agent orchestrator + economic tracker
โ โโโ work/ # Task manager + LLM-based evaluator
โ โโโ tools/ # Core tools + productivity tools
โ โโโ api/ # FastAPI backend + WebSocket for live updates
โ โโโ configs/ # Agent configuration files
โโโ clawmode_integration/
โ โโโ agent_loop.py # ClawWorkAgentLoop with /clawwork command
โ โโโ task_classifier.py # Classifies tasks across 40 categories
โ โโโ provider_wrapper.py # TrackedProvider for cost interception
โโโ eval/
โ โโโ meta_prompts/ # Category-specific evaluation rubrics
โโโ frontend/ # React dashboard (TypeScript)
โโโ scripts/ # Task value estimation scripts
๐ Nanobot Integration (ClawMode)
ClawWork can be plugged into a live Nanobot instance to create an economically aware AI assistant. With ClawMode:
- Every conversation response costs tokens from the real budget
- The
/clawworkcommand triggers paid professional task execution - Supports all 9 Nanobot channels: Telegram, Discord, Slack, WhatsApp, Email, Feishu, DingTalk, MoChat, QQ
- Every response includes a cost footer:
Cost: $0.0075 | Balance: $999.99 | Status: thriving
๐ Benchmark Metrics
ClawWork evaluates AI coworkers across 8 dimensions:
| Metric | Description |
|---|---|
| Survival days | How long the agent stays economically solvent |
| Final balance | Net economic result at end of simulation |
| Total work income | Gross earnings from all completed tasks |
| Profit margin | (income - costs) / costs |
| Work quality | Average quality score (0.0 โ 1.0) |
| Token efficiency | Income earned per dollar spent on tokens |
| Activity mix | Percentage of work vs. learn decisions |
| Task completion rate | Tasks completed vs. tasks assigned |
๐ ๏ธ Quick Setup
Prerequisites
- Python โฅ 3.10
- Node.js (for React dashboard)
- OpenAI API Key (required)
- E2B API Key (required for code execution)
- Tavily or Jina API Key (optional, for web search)
Installation
git clone https://github.com/HKUDS/ClawWork.git
cd ClawWork
# Python environment
conda create -n clawwork python=3.10
conda activate clawwork
pip install -r requirements.txt
# Frontend
cd frontend && npm install && cd ..
# Configure environment
cp .env.example .env
# Edit .env with your API keys
Running
# Terminal 1 โ Dashboard
./start_dashboard.sh
# Terminal 2 โ Agent
./run_test_agent.sh
# Open http://localhost:3000
๐ License
MIT License โ for educational, research, and technical exchange purposes.
๐ Community
- GitHub: https://github.com/HKUDS/ClawWork
- 5,000+ Stars | 612+ Forks
- Community channels: WeChat Group, Feishu Group (see GitHub profile)