ComputerWorld
Computers, phones and an internet, simulated in one deterministic process — for training and testing agents that use a screen.
One engine, and what it is for.
Simulated computers and internet
macOS, Windows, Ubuntu, iOS and Android share one simulated internet. No request leaves the process, and every run replays to the byte.
Expert procedures
Professionals’ procedures performed through pointer and keyboard, then judged by checks on the world’s own state.
Forking and RL
Snapshot a world mid-run and fork it. The same actions from the same checkpoint always reach the same state, so rollouts branch cheaply.
Societies and market research
Many people in one world, each on their own phone or computer. Fork it, change one thing for half of them, and compare.
Hosted API and SDKs
Worlds run on our machines at computaco.ai/v1. The Python and JavaScript clients are MIT-licensed and name their calls after the engine’s own, so code moves between a local world and a hosted one.
from computerworld_cloud import Client
client = Client()
with client.create_world("agent-desktop", seed=1) as world:
env = world.environment()
env.click(interaction="shell:launch:editor")
env.click(role="textbox", label="Document text")
env.type_text("Hello")
env.render().save("screen.png")
fork = world.fork()
run = fork.run("agent-desktop/legal-engagement", seed=1)
print(run["verification"]["passed"])
fork.delete()Pricing
Every plan runs the same engine, the same worlds and the same API.