OpenReward Python SDK

The official Python SDK for OpenReward — a platform for building, hosting, and training on RL environments for language models.

The SDK has two complementary roles:

• Build environments — define evaluation tasks, expose tools, and serve them via a standards-compliant API that can be deployed on the OpenReward platform.
• Train agents — connect to any environment (local or hosted), run agent loops, and log rollouts with rewards back to OpenReward.

Installation

pip install openreward

For environments that process documents (PDF, DOCX, Excel, PowerPoint):

pip install "openreward[tools]"

Requires Python 3.11+.

Core concepts

Environment

An Environment subclass defines a benchmark or task distribution. Implement three required methods:

Actions are defined as async methods decorated with @tool. Each tool receives a Pydantic model as input and returns a ToolOutput.

Calls to the same tool run one at a time per session (different tools still overlap), so parallel calls (e.g. from a multi-agent harness) cannot race a grader or submit twice; once a call returns finished=True, queued calls are refused without running. Mark tools that never grant reward with @tool(concurrent=True) to let them run in parallel. The built-in shell, file and web toolsets already do.

ToolOutput

Every tool returns a ToolOutput containing:

• blocks — a list of TextBlock or ImageBlock results
• reward — optional float reward signal
• finished — whether the episode is complete
• metadata — optional arbitrary metadata

Server

Server wraps one or more Environment classes in a FastAPI app and exposes the Open Reward Standard API over HTTP with SSE streaming.

Key endpoints:

Sandboxes

Environments that need isolated compute (e.g. code execution) can spin up Docker containers via the sandbox API using SandboxSettings. Containers are managed automatically — started in setup() and torn down in teardown().

Overriding session env vars

When calling a hosted environment, pass env_overrides to Environment.session(...) to point the session at a custom inference backend. Owners of the environment may override any env var; other callers are restricted to OPENAI_BASE_URL, OPENAI_API_KEY, ANTHROPIC_BASE_URL, ANTHROPIC_API_KEY — any other key returns a 400.

env = client.environments.get("GeneralReasoning/counter")
session = env.session(
    env_overrides={"OPENAI_BASE_URL": "https://my-vllm.example.com/v1"},
)

Toolsets

Group reusable tools into Toolset classes and compose them across environments via the toolsets class attribute.

Rollout logging

Log agent trajectories with reward signals back to OpenReward for analysis and training. The client's rollout API supports normalized message types as well as raw outputs from Anthropic, OpenAI, and Google GenAI SDKs.

Task difficulty

Environment.get_task_difficulty() returns per-task reward aggregates computed from logged rollouts, grouped by (split, variant, task_index) — useful for spotting which tasks are hardest or have the most variance.

env = client.environments.get("owner/my-env")
for task in env.get_task_difficulty(split="test"):
    print(task.task_index, task.avg_reward, task.num_rollouts)

Each result is a TaskDifficulty with task_index, avg_reward, min_reward, max_reward, num_rollouts, split, and variant. Optional filters: split, model_name, min_model_params, max_model_params, training_stage. If the environment was obtained with a variant, results are scoped to that variant (mirroring list_tasks); otherwise rows span all variants. Data is served from a materialized view, so freshness depends on its server-side refresh.

CLI

The orwd CLI helps you scaffold and create environments.

Scaffold a new environment locally

# Minimal environment
orwd init my-env

# Environment with a Docker sandbox for code execution
orwd init my-env --template sandbox

Create an environment on OpenReward

Registers a new environment under your account (requires OPENREWARD_API_KEY):

orwd create my-env --description "A short description of my environment"

By default the environment is created under your personal namespace. To create it under an organisation you are a member of, pass --namespace:

orwd create my-env --description "A short description" --namespace my-org

Pass --private to make the environment private:

orwd create my-env --description "A short description" --private

Deploying to OpenReward

1. Push your environment to a GitHub repository.
2. Connect the repository in the OpenReward dashboard.
3. Configure compute resources (CPU, memory, scaling).
4. Every push to the connected branch triggers an automatic build and deployment.

Your environment is then accessible to any agent via the OpenReward API using the username/environment-name namespace.

Environment variables

Documentation

Full documentation, guides, and examples are at docs.openreward.ai.

License

Apache 2.0