PlatformBenchmarksModelsComparisonAboutOpen app
A Puma experiment

A Puma experiment

Evals that stay on your machine.

Private suites run in this browser. The public board is a browser-agent harness, and the numbers on it are baselines.

Published baselines · heuristic-scripted 7/11 · naive-first-match 1/11

Private by design

Private evals tell you whether your prompts, tools, and agents hold up on the tasks you actually run. Your cases stay on your device.

Cases you keep

Capture the tickets, documents, and workflows your system must handle. Nothing is uploaded. The cases stay on this device.

Graders you control

Score with exact match, contains, regex, JSON field checks, and keyword lists. Start simple, then tighten the rubric as the work gets clearer.

Runs that stay local

Paste model outputs, label the system under test, and keep a history of pass rates as prompts and models change — all in this browser.

Browser agents, scored locally

11 deterministic tasks on fixture sites that stay on your machine: extraction, navigation, reasoning, forms, and a multi-step login. Checkers score every run. The numbers below are baselines from the open-source harness.

Open the browser-agent leaderboard

Two paths. One purpose.

getEvals sits next to Puma Browser and the PumaAI lab. Same purpose: AI that stays in your hands.

Start with the demo suite Browser agent baselines See the benches