AI playtesting for game studios
Parallel playtesting on demand for game developers
Upload your game and agents test every edge case, finding game-breaking bugs, performance issues, softlocks, broken tutorials, and regressions.
HOW IT WORKS
Connect your dev build
minecraft-1.21-dev.zip
612 MB · uploaded 2 h ago
Send agents to test your game
A goal in plain English — anything you can ask in a sentence. No scripts, no test code.
NEW TESTS
Survive the first night without dying
Craft a stone pickaxe from a fresh spawn
Reach a village before nightfall
Mine iron and smelt an ingot
Agents fan out and play
Every goal becomes a test that runs in parallel on every new build — Keyer keeps score.
Minecraft
survival crafting · build #214 · uploaded 2 h ago
151
runs on this build
5 / 7
goals passing
12
issues found
23 min
avg run length
| Goal | Agent runs | Game-breaking bugs | Performance issues | Warnings | |
|---|---|---|---|---|---|
| Survive the first night without dying | 38 | — | 1 | 2 | View report |
| Craft a stone pickaxe from a fresh spawn | 31 | — | — | 1 | View report |
| Reach a village before nightfall | 22 | 1 | 1 | 3 | View report |
| Mine iron and smelt an ingot | 11 | — | 1 | 2 | View report |
| Craft a furnace | 19 | 2 | 1 | 4 | View report |
| Trade with a villager before nightfall | 16 | 1 | 1 | 5 | View report |
| Reach the Nether with iron gear | 14 | 1 | 2 | 3 | View report |
Every run: video · per-tick trace · referee verdict
Get the report
Bugs, crashes, performance and regressions from every suite — each one tied to a replay you can watch.
Minecraft / Reports /
Build #214
5 / 7
goals passing
12
bugs found
3
new in this build
1
crash caught
58 h
of play, overnight
BUGS FOUND
Placed furnace vanishes after reloading the world save
Craft a furnace · 3 runs · Video · Trace
Agent falls through an unloaded chunk at the cave entrance
Craft a furnace · 2 runs · Video · Trace
Villager trade window can't be reopened after closing it with Esc
Trade with a villager · 5 runs · Video · Trace
Furnace recipe doesn't appear until the inventory is closed and reopened
Craft a furnace · 4 runs · Video · Trace
Respawn screen flashes before the death animation finishes
Survive the first night · 1 runs · Video · Trace
⋯View all 12 bugs →
PERFORMANCE
Avg frame rate
141 fps
1% low
58 fps
Peak memory
3.2 GB
World load
11.9 s
⚠ +3.8 s vs #213
⚠ Leak suspect: +380 MB / hour in the 4 h soak run · trace ↗
CRASHES
Out of memory, 3 h into the soak run
1 of 151 runs · seed 8841 · Minidump · Video · Trace
Every line links to its replay — video · per-tick trace · referee verdict · JSON export
Relive and inspect any error
Watch what the agent played, read what it was thinking — or jump in and take the controls.
Minecraft / Craft a furnace / Run #15 /
Error · HighPlaced furnace vanishes after reloading the world save
OBJECTIVE: Craft a furnace● 12:41 / 23:05+▶MODEL THOUGHTS
Eight cobblestone in the inventory — that's enough. Heading back to the crafting table.
Crafting table open. I place the cobblestone in a ring and the furnace recipe lights up.
Furnace crafted and placed. Referee confirms the goal from the inventory state.
Product dashboard, illustrative data. The demo runs are real →
See the agents in action
Three games. One harness. Zero edits. Plug and play.
RUNS ON BUILDS FROM EVERY MAJOR ENGINE
Unity
Unreal Engine
Godot
GameMaker
CryEngine
+ others
The coverage gap
Every build changes. Your playtests don't keep up.
Human playtests are slow and expensive.
Scheduling players, running sessions, writing up notes — by the time the report lands, the build it describes is three versions old.
Scripted bots test what you told them to.
Recorded input sequences shatter the moment a menu moves or a recipe changes. They can confirm a path still exists — they can't discover that it broke.
The bugs that matter live in the messy middle.
Not in the happy path and not in the unit tests — in the improvised, mistake-riddled way real players actually reach your content.
Capabilities
Built like an experiment, not a demo.
Pixels only
The agent gets rendered frames and the same controls a player has. No API hooks, no source access, no scripted paths. If a human can play the build, Keyer can test it.
Deterministic replays
Seeded worlds and tick-stepped simulation. A failure at tick 512 replays as a failure at tick 512 — on your machine, tomorrow, in CI.
Full evidence
Every run produces a captioned video of what the agent saw and thought, a per-tick input trace, every planning brief, and a machine-readable run report.
Private pilot
Start with one build and one question.
Can a new player reach the first crafting bench? Does the tutorial survive the UI refactor? What breaks when the recipe tree changes? Send us the build and the question — you get back the video, the trace, and the verdict.
FAQ
Fair questions.
Do you need our source code?+
No. The agent interacts with a build the way a player does — rendered frames in, keyboard and mouse events out. The referee verifies outcomes the same way, from the game's own visible state — no SDK, no engine hook, no integration.
Is this replacing our QA team?+
No. It takes the part of QA nobody wants: running the same progression for the two-hundredth time, overnight, on every build. Humans keep the judgment-heavy work — what feels confusing, what feels unfair, what feels fun.
What kinds of games work?+
The harness is engine-agnostic — the three demo runs are a survival-crafting world, a commercial tower defense, and an arcade racer, with no game-specific code between them. It works best on games with deterministic or seedable starts and goals you can state in a sentence: 'reach the first bench', 'survive to round 60', 'finish in under a minute'.
How autonomous is it really?+
Every run on the demo page is unedited and fully autonomous: a planner writes one small brief at a time, an actor executes it from pixels, and a referee checks the result. Nobody touched the controls. Some briefs fail — those stay in the run report too — and the planner routes around them. When a run ends without a win, it's reported as exactly that.
What do we get after a run?+
A verdict verified against game state, the full video with the agent's reasoning captioned over it, the per-tick trace, every brief the planner wrote, and the facts the run added to memory. Everything is diffable between builds.