reacher-z/ClawBench — Open-source benchmark for browser AI agents on 153 everyday online tasks across 144 live websites.

Description and use

Reacher‑z’s **ClawBench** is an open‑source benchmark that evaluates browser‑based AI agents across 153 everyday online tasks drawn from 144 live websites. It combines a five‑layer interaction recording system, DOM‑structure matching, and an LLM‑based judge to score agents, with the current best performance reaching 33.3 %. The project targets AI/ML researchers and data practitioners interested in web‑automation and agent evaluation.

Problem

reacher-z/ClawBench helps remove repetitive manual operations from a workflow.

Outcome

Early or unclear: treat as research material until maintenance, releases, docs, and issue activity are verified.

Workflow

• remove manual work
• connect tools into repeatable flows
• schedule operational tasks

Evidence

Open-source benchmark for browser AI agents on 153 everyday online tasks across 144 live websites. 5-layer recording + DOM-match + LLM judge. Top score 33.3%.

Source