Meta and Hugging Face launch OpenEnv to advance open-source agentic development
New framework provides shared environments for developers building next-generation AI agents.
Meta has announced the launch of OpenEnv, an open-source framework and community hub developed with Hugging Face and a network of AI startups and labs. The announcement was shared by Zach Wentz, Software Engineering Manager at Meta’s Superintelligence Lab, in a LinkedIn post.
A new standard for agentic environments
OpenEnv aims to give developers a shared space to create, package, and share “frontier-grade” reinforcement learning (RL) environments for both training and deployment. The framework builds on infrastructure designed over the past year to support Meta’s AI teams, including FAIR, GenAI, and MSL.
In his post, Wentz wrote: “Excited to share a little something we've been working on at Meta with Hugging Face + many other AI startups and labs: OpenEnv — frontier-grade RL environments for the open-source community.”
He added that the next phase will “allow anyone to create, package, and share environments on the Hugging Face Env Hub and keep working on the spec together with the community.”
Building a shared ecosystem for developers
Meta describes OpenEnv as a system that defines the boundaries and capabilities of agentic environments — the structured spaces in which AI agents operate. These environments specify everything required for a task, including tools, APIs, and credentials, within a secure and sandboxed setup.
According to Meta, the approach offers “clarity, safety, and sandboxed control to agent behavior,” supporting both experimentation and deployment across scalable compute infrastructure.
The OpenEnv Hub will enable developers to interact with and test environments directly, enlist models to perform tasks, and review the tools each environment exposes before large-scale RL training. The framework’s 0.1 specification is now available for community feedback through public RFCs, with integrations under development for TRL, SkyRL, and Unsloth.
Collaboration and next steps
Wentz and the Meta team confirmed that OpenEnv is being integrated with TorchForge, Meta’s new reinforcement learning library, as part of a broader effort to establish shared post-training standards.
The initiative reflects a growing shift among major tech companies toward opening AI infrastructure and tools to the wider developer community. OpenEnv’s collaborative model signals a move to make agentic systems more accessible — and more accountable — through transparent ecosystem design.
Wentz ended his post with a call to the developer community: “Let’s build the future of open agents together, one environment at a time.”
