Author an Environment

Start with EnvFactory when authoring an environment. The same class constructs environments for local evaluations, serving, task sweeps, and managed containers. Put construction in make(), optional one-time setup in prepare(), and adapter roles in tags. If you only need to expose an existing environment object, serve it directly with EnvServer.

A minimal factory

Implement make() and return a Gymnasium-style environment. Keep heavyweight imports inside the method so RLMesh can inspect the factory without loading the simulator.

import rlmesh


class CartPole(rlmesh.EnvFactory):
    def make(self):
        import gymnasium as gym

        return gym.make("CartPole-v1")


if __name__ == "__main__":
    CartPole().serve("127.0.0.1:5555")

Use the same factory in a local evaluation (model.run(CartPole(), episodes=3)) or serve it by import path (python -m rlmesh.serve --env environment:CartPole, with the class saved in environment.py). This is the class structure used by catalog environments.

serve() builds the environment and blocks while clients use the endpoint. A client in another process connects with RemoteEnv and runs the usual reset and step loop. See Connect a Client for that side.

Add a model-facing contract when needed

The factory’s optional tags attribute names the roles of observation and action fields. Declare it when a model needs RLMesh to adapt environment-specific keys, layouts, or encodings to its own input. The factory attaches those tags to every environment it builds; a model’s ModelSpec can then resolve against them. The Adapters guide shows both declarations together.

A factory can also declare construction parameters, enumerate named task variants, and publish a description for a catalog without starting the simulator. The Environment Reference owns the exact hooks, parameter and variant rules, describe() format, and container entrypoint. For a complete container example, see Bring Your Own Container.