OpenAI Open-Sources the Core Framework Behind Codex

OpenAI has taken a significant step toward making AI agents more accessible to developers by fully open-sourcing Harness, the core execution framework that powers its Codex coding agent. Released under the permissive Apache 2.0 license, the framework can be modified and incorporated into other software products, giving developers access to technology that was previously closely associated with OpenAI’s own agent ecosystem.

Harness is important because it provides much more than basic code generation. It handles the agent loop—the process through which an AI system understands a task, maintains context, calls tools, executes actions and works with human approval. The open-source release includes the Codex command-line interface, Codex SDK and app-server components, allowing developers to build their own agent-powered applications around the framework.

The move could significantly expand the use of AI agents beyond traditional chat interfaces. Developers can potentially integrate the framework into engineering platforms, internal business tools and other software where an AI system needs to perform multi-step tasks. OpenAI’s own research shows that enterprise customers are increasingly using Codex for longer and more complex workflows rather than simply asking AI questions.

Open-sourcing the framework also puts OpenAI in a stronger position to benefit from developer contributions and experimentation. Instead of keeping every part of the agent infrastructure proprietary, the company can allow the wider developer community to improve, customize and adapt the technology. Codex itself already has an open-source foundation, with its CLI available under the Apache 2.0 license.

The decision comes during an increasingly competitive AI-agent market, with companies such as Anthropic and DeepSeek also developing tools that allow AI models to operate software autonomously. By releasing Codex’s underlying agent framework, OpenAI is effectively betting that the future of AI competition will depend not only on the intelligence of models, but also on the software infrastructure that allows those models to take action.