Theme
AI Resources
ClawWork
ClawWork is an open-source economic benchmark and nanobot integration from HKUDS. It assigns AI agents professional tasks, tracks model and search costs, evaluates the submitted work, and keeps a running balance that connects output quality with operating cost.
The standalone benchmark uses 220 tasks from OpenAI's GDPVal dataset across 44 occupations. Its ClawMode path can also add cost tracking and an on-demand work command to a nanobot setup, with results and balance changes visible in a local dashboard. Use this as a first read, not a recommendation. Open the original project before trusting details like terms, limits, privacy, cost, setup, or safety.
What it is
A work-and-cost benchmark for AI agents
Agents receive professional tasks, produce artifacts, pay simulated operating costs for model and search use, and receive a quality-based benchmark payment after an automated evaluation.
Why it stands out
Quality and cost share one scoreboard
The dashboard keeps task completions, evaluation scores, token costs, income, balance, and survival status together. That makes an expensive high-scoring run and a cheaper uneven run easier to examine side by side.
Availability
Local source with two setup paths
The repository documents a standalone benchmark with a React dashboard and a ClawMode integration for nanobot. Both require local setup and model credentials; optional search and sandbox providers may need additional keys.
Why it matters
What makes it useful
ClawWork keeps the work artifact, evaluation, token use, search cost, and changing balance in the same experiment, so builders can see whether stronger output also required a much more expensive route.
What to know
Where it fits
Use it as a research and engineering test bed for comparing agent setups on document, spreadsheet, PDF, analysis, planning, and technical tasks. The ClawMode integration is a separate path for adding the same economic accounting to a nanobot gateway.
Notable points
What stands out
The dollar balances and hourly rates are benchmark accounting built from estimated task hours, occupational wage data, automated quality scores, and recorded API costs. They should not be read as customer revenue, employment income, or proof that the same performance will carry into unscripted work.
Before using
What to review
Choose the standalone benchmark or the nanobot integration first; their Python requirements, commands, and configuration paths differ.
Budget for the model calls used by both the working agent and the automated evaluator. Search, hosted code sandboxes, and optional OCR paths can add separate services or credentials.
Treat the LLM evaluator and wage-based payment formula as parts of the experiment. Review submitted artifacts and task-level records instead of relying on the balance or average-quality figure alone.
Use test credentials and a controlled workspace until the selected model, search provider, sandbox, channel, and file tools have the access boundaries you expect.
Check the current repository and MIT license terms before adapting the project or depending on its still-evolving integration path.
Reader fit
Who may find it relevant
Builders comparing agent quality and operating cost on the same professional-task set.
Researchers who want task prompts, submitted artifacts, evaluation records, and cost traces rather than a single final rank.
nanobot users exploring per-message cost tracking and an on-demand professional-task workflow.
Less relevant for people looking for a finished consumer assistant or evidence of real-world wages and customer demand.
Editorial note
Why LifeHubber lists it
ClawWork makes the tradeoff between agent output and operating cost visible at the task level. Readers can judge what produced a high balance, what the evaluator rewarded, and whether the result still looks useful when the work artifact and API bill sit beside the score.
Source links
Source materials
Reader note
Before relying on this entry
LifeHubber lists entries to help readers inspect AI projects, not to endorse them or prove they are safe, suitable, accurate, maintained, or right for a specific use. We do not verify every entry in depth. Before relying on anything listed, review the original materials, terms, privacy practices, limits, and risks that matter for your situation.
What to explore next
Separate the benchmark, the agent, and the score.
Compare broader evaluation results, inspect the agent system being scored, or open nanobot to see the runtime underneath ClawMode.
More in AI Agents
Keep browsing this category
Explore more AI agent projects.
RAGFlow
infiniflow/ragflow
A self-hostable RAG and agent-context platform whose v0.27 line adds Knowledge Compilation, Agentic RAG thinking modes, document and data-source connectors, citations, APIs, and agent workflows.
Auto-Company
MaxMiksa/Auto-Company
A continuous local AI-agent loop that uses Claude Code or Codex CLI to coordinate 14 business and engineering roles, carry decisions through one shared consensus file, and expose runs through a local dashboard.
iFixAi
ifixai-ai/iFixAI
A Python diagnostic for running fixed behavioral and governance inspections against language models and AI agents, with provider adapters, repeatable suites, cross-provider judging, manifests, scorecards, and CLI, CI, plugin, or agent-skill paths.
For project maintainers
Listed here? You can use the badge.
If you maintain a project with a current LifeHubber listing, you may add the optional “Listed on LifeHubber AI Resources” badge to its README, docs, or website. No introduction or permission request is needed.
