For developers
An engine your agent can actually play.
There is no editor to bridge into. The engine is the agent's tool server, and every session plays its own game before it reports anything as done.
-
built Agents that play and prove their own work
Each session steps the world, sends player input, walks the player to a point, queries what happened and pins it with a test. Code that “appears valid but the game is actually unplayable” is caught before anyone reads it.
-
built Many sessions at once, without stepping on each other
A private headless engine per session, focused on one feature folder, so ten agents build ten slices in parallel. Rebuilding the engine never blocks a running session.
Starting the sessions from one request is not built yet; today you start them yourself.
-
built Deterministic replay
Snapshots, restore and replayable play sessions turn “it happened once” into a test.
-
built Ask why instead of adding logging
Walk a system line by line with the values its variables took, trace fields over time, and list systems in tick order with what they read and write.
-
built Everything is text
Scripts, generators, scenes and reports diff and merge like code, so agents and code review both work on them.
-
built Typed scripting with instant answers
TypeScript, a definite typecheck answer for what is on disk, and hot reload that reports the live version.
-
built Bring your own agent
MCP over standard input and output, no port or editor needed: Claude Code, Codex, Cursor or KiwiPow IDE.
-
built A way out for hot paths
Move a system to WebAssembly or a native bundle when a script is too slow, measured first.
-
built Reusable bundles
Installable feature packs, such as walking, that an agent learns from their guide.
Few bundles exist yet.