Benchmarks AI agent browser automation against PinchTab across chrome, cloak, and ghost-chrome providers.
npx skills add https://github.com/pinchtab/pinchtab --skill pinchtab-opt
Please help me install this Agent Skill. Skill: pinchtab-opt Source: https://github.com/pinchtab/pinchtab/tree/main/skills/pinchtab-opt Command: npx skills add https://github.com/pinchtab/pinchtab --skill pinchtab-opt
Measuring how efficiently an AI agent can drive a browser through PinchTab requires a repeatable, instrumented benchmark; this Skill runs a standardized 108-step optimization loop and a fresh-install setup validation so doc quality and agent performance are quantified instead of guessed.
all argument./pinchtab-opt setup all to confirm a weak model can still complete the 11-step fresh-install flow on every provider.Ask the agent to run the PinchTab optimization loop against the chrome provider, or run the setup test with the setup keyword followed by a provider name.
High-intent search queries and answers about installing and using this skill.
Invoke /pinchtab-opt with an optional provider argument: chrome (default), cloak, ghost-chrome, or all. The skill brings up the provider with up.sh, spawns three parallel subagents over 47 task groups, then merges their JSON reports and prints a baseline comparison table.
Optimization mode runs three Docker-based subagents on the runner's HIGH model across 108 steps to measure browser-op efficiency. Setup mode runs one native-binary subagent on the LOW model to test whether the skill docs alone enable a zero-to-working fresh install.
The benchmark supports chrome, cloak, and ghost-chrome providers. Passing all runs each provider sequentially, and up.sh accepts all three names while down.sh tears down the shared containers.
The setup test is a documentation-quality gate: if a small, fast model like Claude Haiku can complete all 11 steps from the skill docs alone, the onboarding flow is genuinely out-of-the-box ready. A strong model passing would not prove the docs are sufficient.
The stored baseline completes 108 of 108 steps using 272 browser operations in about 49.8 seconds, averaging 2.5 ops per step. Agents are expected to land in the 250-400 ops range, roughly 2.5-4 ops per step.
i0switchManage a persistent local Chrome service for shared browser automation sessions.
kis9aExecute Go benchmarks and analyze ns/op metrics to identify bottlenecks.
sumik5Automate web tasks via agent-browser CLI controlling Chrome through CDP.
SinesysTechDevelop data tables and dashboards with TanStack Table, shadcn/ui, and Recharts.