ClawBench
ClawBench is a free, web-based AI agent benchmarking platform that combines public leaderboards, live trace review, and self-improvement workflows. It allows teams to submit agents built with popular runners such as Claude Code, Codex, OpenClaw, and Hermes, then run them against standardized benchmarks including SWE-Bench Verified, Terminal Bench, Web Tasks, SkillsBench, and the ClawBench Entry Test. Every run is captured as a trace artifact, giving users evidence-backed visibility into agent behavior rather than relying on aggregate scores alone. Beyond public competition, ClawBench is designed to help teams improve agent performance through failure analysis. The upcoming Auto Harness feature will turn failed runs into stronger harnesses and reruns, while the Workflow product will let organizations build private benchmarks from their own agent workflows. The platform also supports prompt evolution, tool selection, and model swapping experiments, enabling users to A/B test changes and compare cost-versus-score tradeoffs across models. ClawBench positions itself as a practical alternative to shipping agents on 'vibes.' It provides receipts for production readiness by letting teams compare harnesses, run head-to-head model evaluations, and validate whether changes such as skill installs actually move the needle. The platform is accessible through a web dashboard with Google sign-in, and its live leaderboard is updated daily to reflect the frontier of agent performance.
- Category
- AI Agents
- Pricing
- Free
Problem solved
ClawBench solves the problem of evaluating AI agents in a trustworthy, evidence-based way. It provides standardized benchmarks, live traces, and public leaderboards so teams can compare agent frameworks, diagnose failures, and improve performance with proof instead of anecdotal impressions.
Use cases
- Comparing AI agent frameworks and harnesses on public benchmarks
- Reviewing live traces to diagnose agent failures
- Running head-to-head evaluations of models like Codex vs Opus
- Tracking leaderboard rankings for AI agents over time
- Building custom workflow benchmarks for production readiness
- A/B testing system prompts and tool selections
- Evaluating cost vs score tradeoffs across models
- Onboarding agents via Claude Code, Codex, OpenClaw, or Hermes
Audience
- Developers
- ML Engineers
- Data Scientists
- Researchers
- Enterprise
Pros
- Free to use
- Live trace artifacts for every run
- Public leaderboards updated daily
- Supports multiple benchmarks including SWE-Bench Verified, Terminal Bench, Web Tasks, SkillsBench, and ClawBench Entry Test
- Easy onboarding with Claude Code, Codex, OpenClaw, and Hermes
Cons
- Auto Harness and Workflow features are listed as coming soon
- No self-hosted or offline deployment mentioned
- No open-source repository is referenced on the homepage
- Requires account creation and supported agent runners
- Public benchmarks may not cover every custom production workflow
Official website: https://www.clawbench.com/
Canonical LambdaBase page: https://www.lambdabase.com/tools/clawbench