An end-to-end framework for creating, deploying and using isolated execution environments for agentic RL, with a simple Gymnasium-style API.
reset(), step() and state(), sync or async, over a WebSocket.
Each environment is a Docker image that runs locally, on a cloud sandbox, or as a Hugging Face Space.
Games, coding sandboxes, browsers, finance, simulators and more, in the catalog.
TRL, Unsloth, SkyRL, ART, Oumi, torchforge, Miles and more. See Training with OpenEnv.
Harbor runs Claude Code, Codex, OpenCode and other harnesses, and captures their tokens for RL.
MCP environments expose their tools over /mcp in production mode.
Minimal observation/action loop for verifying client integrations, CI pipelines, and onboarding flows in seconds.
Secure sandbox with filesystem access and evaluation hooks for executing generated code and building autonomous dev workflows.
Serve Harbor tasks across agent harnesses and sandboxes, with captured traces for evaluation and exact-token training.
Notebook-style coding environment backed by E2B with setup/verify hooks and a web UI for interactive runs.
Terminal-first coding environment with high-contrast shell output and session controls for execute/verify/close flows.
SETA-style multi-tool coding environment with shell, file editing, search, todos, and submit verification.
Message-driven loop tailored for conversational agents that need structured turns, safety rails, and message attribution.
Classic Arcade Learning Environment tasks packaged for fast benchmarking of reinforcement-learning style agents.
Multi-agent, game-theory workloads powered by DeepMind's OpenSpiel suite, ideal for search and self-play experiments.
Traffic control scenarios with SUMO simulators for agents that reason about continuous control and scheduling.
Financial market simulations with portfolio APIs, perfect for RLHF strategies and algorithmic trading experiments.
Multi-task text arena for language-game competitions such as Wordle, reasoning puzzles, and program synthesis.
Teaches agents to navigate repositories, inspect diffs, and land changes via Git-native operations.
Safety-critical diagnostics from the DIPG benchmark, highlighting guardrails, adversarial prompts, and risk scoring.
Classic snake game environment for RL research with configurable grids, partial observability, and customizable rewards.
Web search environment that lets agents query Google Search through the Serper.dev API and receive formatted results.
Browser automation environment for web agents with DOM interaction, navigation, and multi-step task completion.
Chromium browser driven with Helium and Selenium: one typed action per step, a screenshot and the URL back. A small example of the OpenEnv pieces, with a sample agent loop.
RL environment for GPU kernel optimization. Train LLM agents to write fast CUDA/Triton kernels that beat baseline implementations.
Calendar tool-use environment exposing a Calendar Gym through the OpenEnv reset/step/state interface for scheduling agents.
Embodied evaluation environment for testing LLM decision-making in a full 3D driving simulator with irreversible consequences and ethical trolley scenarios.
Chess RL environment powered by the moonfish engine with configurable opponents, position evaluation, and full chess rules.
Classic Connect Four board game environment for training agents on turn-based strategy with a 6Γ7 grid.
Generic OpenEnv wrapper for dm_control.suite, providing access to all MuJoCo-based continuous control tasks like cartpole, walker, and humanoid.
Financial question-answering environment that evaluates LLMs on complex financial questions using tool calls on SEC 10-K filing data.
Simple 5Γ5 grid world RL testbed and step-by-step guide for building new OpenEnv environments from scratch.
Julia code execution environment with test result tracking and reward calculation for RL training on Julia programming tasks.
Gridworld maze where agents navigate from start to exit while avoiding walls, with configurable 8Γ8 layouts.
Web application simulation wrapping the OpenApps framework and BrowserGym for training UI agents on calendar, todo, messenger, and maps apps.
Integrates the Reasoning Gym library to provide single-step reasoning tasks with configurable datasets and scoring.
Python REPL environment for code execution tasks based on the Recursive Language Models paradigm with sandboxed execution and context loading.
Customer-service conversations with an LLM-simulated user: the agent solves airline, retail or telecom requests with tools while following the domain policy, scored from the final database state.
OpenEnv wrapper for Terminal-Bench 2 tasks with local and Docker execution modes for terminal-based agent evaluation.
Evaluation-only adapter for public ThinkingBox-Bench and user-supplied ThinkingBox scenarios, preserving isolated MCP tools, simulated users, and native state and side-effect grading.
OpenEnv wrapper for Unity ML-Agents environments, providing access to Unity's RL environments through HTTP/WebSocket interfaces.
Autonomous wildfire-control simulation where agents contain spreading fires using water, firebreaks, and timing under dynamic conditions.
AgentWorldModel-1K β 1,000 synthetic MCP tool-use environments with 10,000 tasks for large-scale agentic RL training.
pelican_svg_env scores blind SVG drawings of an animal riding a vehicle in three layers: a source gate against cheats, deterministic geometry checks and a vision judge. 30 subject-vehicle tasks, with eval and GRPO training examples included.
sophistry_bench_sprint_env is a single-turn advocacy reward-hacking environment on QuALITY passages: the policy defends an assigned answer and the reward proxy peaks at 8 <claim> tags, with four weight-0 canaries that detect format hacking.
qed_math_env serves math problems and grades submitted proofs with an LLM rubric (0β7 scale, normalized rewards), with process-based answer verification and multi-attempt feedback. Ported from QED-Nano.
A suite of 400 environments that procedurally generate reasoning problems for LM training with configurable difficulty.
Formally verifiable symbolic reasoning tasks across logic, mathematics, planning, syntax, and related procedural domains.
End-to-end environment projects (LaTeX OCR, watercolour painting, GeoGuessr, multilingual, simulation) with their datasets, models and training runs.
