OpenEnv is a tool for creating an agent execution environment, such as a terminal, browser, or anything an agent can interact with. And today, we’re excited to announce that OpenEnv is becoming even more open, making the future of agent training open source.
Starting today, OpenEnv will be coordinated by a committee that so far includes Meta-PyTorch, Reflection, Unsloth, Modal, Prime Intellect, Nvidia, Mercor, Fleet AI, Microsoft, and Hugging Face. OpenEnv now resides in huggingface/OpenEnv.
The OpenEnv project is supported and adopted by some of the leading organizations in the AI ecosystem, including PyTorch Foundation, vLLM, SkyRL (UCB), Lightning AI, Axolotl AI, Stanford Scaling Intelligence Lab, Mithril, OpenMined, Scaler AI Labs, Scale AI, Patronus AI, Surge AI, Halluminate, Turing, Scorecard, and Snorkel AI.
Why you need OpenEnv for training open source agents
Agent harnesses such as Claude Code, Codex, OpenClaw, and Hermes continue to improve. One reason for the improvement is that models such as GPT-5.5 and Opus 4.8 are trained to use their respective harnesses.
We want to realize these benefits in open source models as well, such as training local models that use harnesses effectively and saving compute by specializing models for specific tasks.
Why do we need to be more open?
Frontier Labs trains models and harnesses that perform just as well with gloves on, in most cases. The model is trained to use the harness and is optimized for its characteristics. The model can generalize beyond these harnesses to some extent, but nothing beats training efficiency.
Not so in public. Developers use any harness, any model, any inference engine, depending on the use case they care about. This is fundamental to the community, but it is also a challenge that requires infrastructure and tools to tackle.
That’s where OpenEnv comes into play. OpenEnv is a library that interfaces between harnesses, environments, and trainers and works with any model. To sustain this, it must be owned by all key stakeholders.
Protocol layer rather than reward framework
Alongside the governance changes, OpenEnv has also become more stringent.
In recent releases, OpenEnv has become an interoperability layer for RL environments. Its role is to standardize the way agents expose, deploy, and use environments. It does not determine how rewards are defined or how the training loop works. Reward definitions, scoring rubrics, and trainer-specific logic belong to libraries dedicated to them. OpenEnv is a common socket that all users can connect to.
In practice, this means:
One interface, many environments expose familiar gym-style APIs (reset(), step(), state()) that run in a client/server architecture. Trainers that speak OpenEnv can drive compliant environments without the need for custom code.
Familiar protocols and formal packaging. Environments are served via standard protocols such as HTTP and WebSockets and packaged with Docker. Because MCP is a first-class citizen, OpenEnv environments are immediately compatible with MCP servers, and the same environment works consistently in both simulation (training/evaluation) and production modes.
Interoperability between env libraries. You can define and utilize environments across different ecosystems (verifiers, ports, etc.) and on the infrastructure or hub of your choice. OpenEnv is not a competitor to them, but a deployment and interface layer underneath them.
what’s next
Over the next few months, our focus will be on transforming OpenEnv from a fast-growing project to a trusted standard.
External Rewards: You can use OpenEnv as a deployment layer to define rewards in libraries you already use (RFC 006). Tasksets via datasets: Wire environment tasks to Hugging Face datasets to ensure environments and benchmarks are configured cleanly (RFC 007). Continuous Harness Integration: First-class support for agent harnesses. End-to-end examples: Complete training and evaluation walkthroughs for TRL, Unsloth, and more. Automated validation: Measure the quality of the environment and its contribution to model learning. This gives the community a scalable way to assess the environment and improve quality (think hackathons!). RFC008.
participate
OpenEnv has a community-centered design and is still in its early stages. Expect some rough spots and help smooth them out. Check the code and RFC: github.com/huggingface/OpenEnv
Thank you to everyone who helped make this transition possible. Let’s build a common foundation for open source agent RL together.
