What does this skill do, and when should you use it?
ClawWork is an MIT-licensed open-source economic survival benchmark from HKUDS. AI agents start with just $10, pay real-time token costs on every API call, and earn $82–$5,000 per completed task drawn from OpenAI's GDPVal dataset — 220 real professional tasks across 44 occupations — with payment scaled by quality score and BLS wage estimates. The system ships with a React dashboard, an economic tracker, and a multi-model leaderboard where top agents reach roughly $1,500+/hr equivalent earnings. It runs standalone as a simulation or plugs into an existing Nanobot gateway via ClawMode, converting a chatbot into a cost-aware AI coworker.
In each daily cycle the agent calls decide_activity to choose 'work' or 'learn'. When working, it reads a GDPVal task and may use web search, a code sandbox (E2B by default), file creation (txt/xlsx/docx/pdf and more), and video generation tools to produce deliverables, then submits via submit_work for LLM evaluation against category-specific rubrics and payment; learning saves knowledge to persistent memory via learn. A TrackedProvider intercepts every LLM call and deducts token costs; the economic tracker logs balance, income, costs, and survival tier (Thriving/Stable/Struggling/Bankrupt), a FastAPI + React dashboard streams metrics over WebSocket, and accounting is written to local files like token_costs.l.
- AI researchers benchmarking different models (GLM, Kimi, Qwen, Gemini, etc.) on real work quality, cost efficiency, and long-term economic survival
- Engineering teams adding economic accountability to an existing Nanobot gateway spanning nine channels (Telegram, Discord, Slack, etc.) so every conversation carries a cost
- Educators or demo builders wanting to visualize how AI makes work-vs-learn career decisions under budget pressure
- Benchmark developers extending the system with new task sources, evaluation rubrics, or LLM providers using the documented extension points
How do you install this skill?
- Static assessment only; no code was executed. All scores derive from source reading with low confidence.
- The skill is injected via always:true and instructs the agent to perform web searches and sandboxed code execution; understand cost and external-call implications before use.
- Earning figures in the README ($15K/$19K) are self-reported and not independently verifiable from static material.
- Core function depends on overseas services (OpenAI, E2B, Tavily); mainland-China reachability is not addressed.
- The skill file has no version or changelog; updates may change agent behavior without notice.
- Publisher is not verified by the FollowSkills curated registry; identity is treated as unknown.
- Shell / CLI
- Network access
- Local filesystem
Python 3.10+OpenAI API keyNode.js (dashboard frontend)E2B sandbox (default code execution)
Clone and enter: git clone https://github.com/HKUDS/ClawWork.git && cd ClawWork. Create a Python 3.10+ environment (conda or venv) and run pip install -r requirements.txt. For the dashboard, cd frontend && npm install. Copy cp .env.example .env and fill in OPENAI_API_KEY (required); optionally set E2B_API_KEY, WEB_SEARCH_API_KEY, and provider variables. Full ClawMode integration steps are in clawmode_integration/README.md (not reproduced on this page).
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/HKUDS/ClawWork.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/clawmode_integration/skill" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Standalone mode: run ./start_dashboard.sh in one terminal to launch the FastAPI backend and React frontend, and ./run_test_agent.sh in another to start the agent; watch live balance and task completions at http://localhost:3000. Agent configuration lives in JSON files under livebench/configs/, where you set initial balance, token pricing, date ranges, and multiple agents. For ClawMode, run python -m clawmode_integration.cli agent|gateway to add economic tools and the /clawwork command to an existing Nanobot instance; note balance tracking only works through the ClawMode gateway — direct nanobot agent runs bypass the tracker.
What are this skill's strengths and limitations?
- Grounded in real GDPVal tasks with BLS-wage-based pricing tied to quality scores — evaluation dimensions mirror production environments
- Genuine cost accounting: token costs are read from API responses rather than estimated, and economics are transparently logged
- Lightweight architecture — one pip install plus a config file deploys an economically accountable agent, with clearly documented extension points
- Ships with a React dashboard and a public multi-model leaderboard for strong visualization
- Requires your own OpenAI API key and real API spend; the default E2B code sandbox is an external dependency with rate limits
- Agents must submit by iteration 10–12 of 15 or time out, and quality scores below 0.6 earn nothing — failure costs fall on the operator
- No evidence in the docs of an automated test suite; stability must be verified yourself
- Headline earning figures ($15K/11 hours, $19K/8 hours) come from the project's own leaderboard and lack independent verification
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| ClawWork: AI Economic Survival Benchmark & Coworker this page | 43 · Not recommended | ★ 8.6k | 7mo ago | MIT |
| AgentTrace | 57 · Use with care | ★ 145 | 19d ago | MIT |
| OpenMAIC Multi-Agent Interactive Classroom Skill | 65 · Recommended | ★ 41k | 11d ago | MIT |
| Long-Horizon Prompting | 57 · Use with care | ★ 18k | 14d ago | MIT |
| Hugging Face Evaluation Manager Skill | 54 · Use with care | ★ 38k | 13d ago | MIT |
The source positions ClawWork as an alternative to traditional technical benchmarks, which measure only technical metrics; ClawWork additionally measures work quality, cost efficiency, and long-term economic sustainability, built on OpenAI's GDPVal dataset. No other named competitors appear in the source.
How did FollowSkills review this skill?
The SKILL.md is a pure instruction protocol with no direct credential or exfiltration behavior, but it is force-injected via always:true and instructs the agent to use web search and code sandboxes without disclosing data flows, confirmation gates, or rollback; deducted for missing least-privilege scoping and external-effect disclosure.
Instructions are internally consistent (workflow, thresholds, iteration caps), but the skill fully depends on external nanobot/clawwork tools with no described tests, error handling, or failure feedback; deducted because key paths are not reproducible statically and abnormal-input behavior is unknown.
Scenario is clear (economic survival loop in a benchmark), but non-fit boundaries are undeclared, and core function depends on overseas services (OpenAI, E2B, Tavily) with no mainland-China reachability guidance; deducted for thin boundary and environment-fit evidence.
Repository has MIT license, README, troubleshooting, and roadmap, but the skill file itself lacks versioning/changelog, its description is somewhat marketing-flavored, and maintenance ownership is not stated at the skill level; deducted for incomplete versioning and skill-level governance.
As a benchmark simulation prompt it plausibly guides the agent loop, but end-user value depends on running the full ClawWork system; the $19K/$15K earning claims are not verifiable from static files, deducted for limited evidence of direct usability and marginal benefit.
README offers a leaderboard and GDPVal references plus a Pages deploy workflow, but none cover the skill's key paths with reproducible tests; earning figures lack auditable primary material, deducted for single-source, non-reproducible evidence.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →