Build it in any framework. denv runs it as a whole system, with its data, on your cloud or any sandbox. Agents run in it, get evaluated in it, and models train against it.
$ denv deploy alpha · egress denied (verified) · snapshot ok $ denv fork --name rollout-07 alpha forked in 0.9s · data included $ denv verify rollout-07 agent stopped · graded on a fork · reward 1
The framework is yours; the environment is denv’s. Everything is a CLI command, a REST call, and an MCP tool with the same name.
A repo with a compose file. Push to deploy; every deploy is a version with a consistent snapshot.
Your own machines, a cloud account, or a sandbox provider. We never own your compute.
Harbor, Inspect, HUD, or your harness runs inside. A fork takes about a second, data included.
Every rollout records the evidence: digests, the egress probe, the caps, the verifier’s output.
Enforced when the environment is resolved. Grading data never enters the agent’s fork.
Agent frozen, submission copied to a root-owned location, graded on a separate fork. Symlinks, pipes, and oversized files score 0.
Egress denied and proven by a probe before every episode. Memory, processes, and CPU capped per service.
We run the environment. We do not train your model.
Sixteen rollouts of one prompt are sixteen forks of one version, about a second each. GPUs stop waiting on cold starts.
harbor run -e denv.harbor:DenvEnvironment and provider=DenvProvider() reach SkyRL, verl, rLLM, prime-rl, TRL, and Tinker.
denv verify produces it as a separate principal. The environment has a version; a run records which one it trained against.
Thirty minutes. We deploy it in your account, fork it, and grade it. No charge.