Quickstart

Serve a Gymnasium environment in one process. Connect to it from another process.

Install

pip install "rlmesh[gymnasium,numpy]"

See Installation for the optional extras.

Server

Run this in the first process:

import gymnasium as gym
import rlmesh

env = gym.make("CartPole-v1")
rlmesh.EnvServer(env, "127.0.0.1:5555").serve()

Client

Run this in a second process:

from rlmesh.numpy import RemoteEnv

env = RemoteEnv("127.0.0.1:5555")
obs, info = env.reset(seed=0)

terminated = truncated = False
while not (terminated or truncated):
    action = env.action_space.sample()
    obs, reward, terminated, truncated, info = env.step(action)

env.close()

The server owns the Gymnasium environment and its dependencies. The client needs only the endpoint address and the spaces the server reports.

Run a Model Against It

Wrap a prediction function in rlmesh.Model and let it pump whole episodes:

import rlmesh


def predict(observation):
    return 0


model = rlmesh.Model(predict)
result = model.run("127.0.0.1:5555", max_episodes=3)
print(result.mean_reward)

run dials the endpoint, plays three episodes, and returns a RunResult with per-episode rewards, mean_reward, and success_rate.

Runnable Files

From the repository root:

uv run python examples/python/quickstart/serve_gymnasium.py

In another terminal:

uv run python examples/python/quickstart/eval.py

Swap in another Gymnasium registration with --env-id:

uv run python examples/python/quickstart/serve_gymnasium.py --env-id Acrobot-v1

For the smallest custom environment object, use examples/python/quickstart/serve.py. It implements a tiny Gymnasium-style CounterEnv without installing Gymnasium.

Next

• Running Evaluations: run, session, seeds, and reading results.

• Models: wrapping a predict callable or subclassing a backend Model.