DecadaTest
AI-powered game QA

An AI agent that plays your game and files the bugs.

Point the desktop app at a web game and a vision-model agent play-tests it like a human would — exploring menus, trying inputs, flagging bugs — then hands you a reproducible report with recorded video evidence. It runs locally, on your machine.

Free to start · see how the models score →

decada · live session · pixel-dashrunning
step 47 · game paused while the agent thinks
thinkWall-slide works, but the second jump after it never fires. Re-testing…
press_keySpace ×2
wait_frames24 frames
thinkConfirmed — the player fell. Double-jump is consumed by the wall-slide.
flag_bugDouble-jump unavailable after wall-slide · high
noteReproduced 2/2 times on level 2.
● 1 bug·video recording on

How it works

From build to bug report in one session.

01

Point it at your game

Download the desktop app and add a hosted URL or an HTML5 build. itch.io embeds, Godot, Unity, plain canvas — if it runs in a browser, the agent can play it.

02

The agent plays

A vision model drives a real browser — reading the screen, pressing keys, clicking, exploring menus. The game is frozen while it thinks, so even fast-paced gameplay is testable. Watch live and steer it with chat.

03

Get a reproducible report

A verdict, a list of bugs with severity and repro steps, and a recorded video of the whole session — so you can see exactly what broke, and when.

Built for real games

Not a DOM crawler with a prompt. A purpose-built harness that makes vision agents reliable testers.

Autonomous exploration

An orchestrator agent reads your game, maps its controls and menus, and drafts a structured test plan — no scripting, no selectors, no setup.

Deterministic playback

The runner pauses the game loop while the model thinks and advances time frame-by-frame per action. No latency pressure, no flaky timing.

Live co-pilot

Watch the agent play in real time with its reasoning streamed alongside every frame — and redirect it mid-session with a chat message.

Recorded evidence

Every session ships with a full video recording and a step timeline. Bugs come with the moment they happened — not a stack trace to guess from.

Runs on your machine

Sessions run locally in the desktop app — your builds never have to leave your computer. The cloud keeps your dashboard, community, and billing in sync.

Bring your own model

Use the hosted model tiers, or point Pro at any OpenAI-compatible endpoint — vLLM, Ollama, LM Studio, or a cloud provider. Your key never touches our servers.

Free for everyone

Which model actually finds bugs?

Every model on the leaderboard is scored on the same mutation-tested suite: can it find the planted bug in a broken game without crying wolf on a clean one? Run it against any model — including your own — from the desktop app or the CLI, and put your result on the board.

# score any model on the bug-finding suite

$ npx decada-bench

precision 0.91 · recall 0.84 · F1 0.87

✓ submitted to leaderboard

Watch the agent work

Published test sessions from the community — reasoning, bugs, and full recordings.

View all

Loading community runs…

Let an agent take the first pass.

Download the desktop app, add your game, and get your first bug report today — on your own machine, with your choice of model.