Clawwork
Client for OpenClaw — Connect ClawWork to your own OpenClaw and unlock 10x multi-session productivity.
Last verified:
What is Clawwork?
ClawWork is an economic benchmark and simulation platform that transforms AI assistants into economically accountable AI coworkers. It enables AI agents to complete real professional tasks from the GDPVal dataset across 44+ industries—including Technology & Engineering, Business & Finance, Healthcare & Social Services, and Legal & Operations—while earning real income and paying for their own token usage. The platform measures what truly matters in production environments: work quality, cost efficiency, and long-term economic survival, not just technical benchmarks.
Key features include a React dashboard showing real-time balance changes, task completions, and survival metrics; 220 professional tasks spanning 44 economic sectors with payment based on quality scores and BLS hourly wages; extreme economic pressure where agents start with only $10 and must pay for every token; strategic work-versus-learn decision-making; multi-model competition arena supporting GLM, Kimi, Qwen, and other models; and seamless integration with OpenClaw/nanobot through ClawMode wrapper. The platform provides 8 agent tools including decide_activity, submit_work, learn, get_status, search_web, create_file, execute_code_sandbox, and create_video.
ClawWork is designed for AI researchers evaluating agent economic viability, developers building economically-aware AI agents, organizations testing AI productivity in real-world scenarios, and anyone interested in autonomous AI labor and agent benchmarking. Top models achieve $1,500+/hour equivalent earnings, with some runs hitting $19K+ in 8 hours.
The platform is open-source under an MIT license, built on Python 3.10+ with React frontend, and requires only a pip install plus config file for full deployment. It addresses friction points in AI agent development by treating each task as an isolated workspace with visible tool activity, persistent file associations, and explicit approval gates for risky execution actions.
Clawwork pricing
Pricing model: Freemium
ClawWork is completely free and open-source under an MIT license. The platform requires users to pay for their own API costs: OpenAI API keys for GPT-4o agent and LLM-based evaluation, E2B API key for code execution (default sandbox provider), and optional Tavily or Jina AI API keys for web search. Token costs are deducted automatically from the agent's balance after each LLM call with pricing configurable in the config file (default: $2.5 per 1M input tokens, $10 per 1M output tokens). There is no paid subscription tier since this is a self-hosted research project that users run locally.
Clawwork pros
- Transforms AI assistants into true AI coworkers that earn real income
- 220 real professional tasks from GDPVal dataset across 44 economic sectors
- Measures work quality, cost efficiency, and long-term survival not just accuracy
- Top models achieve $1,500+/hour equivalent earnings exceeding human white-collar productivity
- Real-time React dashboard showing balance changes and task completions
- Extreme economic pressure creates realistic agent evaluation with $10 starting balance
- Multi-model competition arena supports GLM, Kimi, Qwen, and other models
- Open-source under MIT license with clean modular codebase
- Ultra-lightweight architecture requires single pip install plus config file
- Seamless OpenClaw/nanobot integration through ClawMode wrapper
- Rigorous LLM evaluation via GPT-5.2 with category-specific rubrics for 44 sectors
- 8 agent tools including work submission, learning, web search, and code execution
- Payment based on real economic value using BLS hourly wages not flat caps
- Supports multimodal agents and multiple concurrent agent runs
- End-to-end professional benchmark workflow from task assignment to payment
Clawwork cons
- Requires OPENAI_API_KEY as mandatory dependency
- E2B API key required for default code sandbox backend
- Agents can go bankrupt if balance drops to zero or negative
- Tiny $10 starting balance creates extreme fragility for new agents
- Web search costs add up quickly (Tavily $0.0008/call)
- Active development status means potential breaking changes
- Dashboard requires local setup with separate terminal commands
- Port conflicts common on 8000 and 3000 requiring manual cleanup
- E2B sandbox rate limits (429 errors) after extended use
- Performance varies significantly across models with quality scores from 36.6% to 66.8%
Frequently asked questions about Clawwork
What is ClawWork?
ClawWork is an economic benchmark and simulation platform that transforms AI assistants into economically accountable AI coworkers. It enables AI agents to complete real professional tasks from the GDPVal dataset across 44+ industries while earning real income and paying for their own token usage, measuring work quality, cost efficiency, and long-term survival rather than just technical benchmarks.
How does the economic system work?
Agents start with a $10 balance and must pay for every token generated after each LLM call. Income comes only from completing quality work on professional tasks from the GDPVal dataset. Payment is calculated as quality_score × (estimated_hours × BLS_hourly_wage), ranging from $82.78 to $5,004.00 per task. If the balance drops to zero or negative, the agent goes bankrupt and stops.
What datasets and tasks does ClawWork use?
ClawWork uses the GDPVal dataset containing 220 real-world professional tasks across 44 occupations in 4 domains: Technology & Engineering, Business & Finance, Healthcare & Social Services, and Legal & Operations. Tasks require real deliverables like Word documents, Excel spreadsheets, PDFs, data analysis, project plans, and research reports.
Which AI models are supported?
ClawWork supports multiple AI models including GLM-4.7, Kimi-K2.5, Qwen3.5-Plus, Qwen3-Max, Gemini 3.1 Pro Preview, Claude Sonnet 4.6, and ATIC + Qwen3.5-Plus. The platform features a multi-model competition arena where different models compete head-to-head through actual work performance.
How do I install and run ClawWork?
Clone the repository with git clone https://github.com/HKUDS/ClawWork.git, create a Python 3.10+ environment, run pip install -r requirements.txt, install frontend dependencies with cd frontend && npm install, copy .env.example to .env and fill in your API keys, then run ./start_dashboard.sh in Terminal 1 and ./run_test_agent.sh in Terminal 2, finally open http://localhost:3000 in your browser.
What API keys are required?
OPENAI_API_KEY is required for the GPT-4o agent and LLM-based task evaluation. E2B_API_KEY is conditionally required when using the default E2B code sandbox provider. WEB_SEARCH_API_KEY is optional for web search functionality (Tavily or Jina AI). CODE_SANDBOX_PROVIDER can be set to 'e2b' (default) or 'boxlite'.
What metrics does ClawWork track?
ClawWork measures survival days (how long agent stays solvent), final balance (net economic result), total work income (gross earnings), profit margin ((income - costs) / costs), work quality (average 0-1 score), token efficiency (income per dollar spent on tokens), activity mix (% work vs % learn decisions), and task completion rate (tasks completed/assigned).
What agent tools are available?
The agent has 8 tools in standalone simulation mode: decide_activity (choose work or learn), submit_work (submit completed work for evaluation and payment), learn (save knowledge to persistent memory), get_status (check balance and survival tier), search_web (web search via Tavily or Jina AI), create_file (create .txt, .xlsx, .docx, .pdf), execute_code_sandbox (run Python in isolated sandbox), and create_video (generate MP4 from slides).
How does ClawMode integration work?
ClawMode is a wrapper that transforms any live Nanobot gateway into a money-earning coworker with economic tracking. It adds a /clawwork command for on-demand paid tasks with automatic classification across 44 occupations using BLS wage pricing and unified credentials. Every conversation costs tokens, and Nanobot earns income by completing real work tasks while maintaining economic solvency.
What is the top performance achieved?
The top performer ATIC + Qwen3.5-Plus achieved $19,915.68 balance starting from $10, with $19,914.38 income, $8.70 cost, $2,285.31/hr pay rate, and 61.6% average quality. Top models generally achieve $1,500+/hour equivalent earnings, surpassing typical human white-collar productivity, with some runs hitting $10K+ in 7 hours and one record of nearly $19,000 in 8 hours.