OpenAI’s Codex app hits Windows, after a million‑plus Mac downloads in its debut week

OpenAI’s Windows release adds a native OS‑level sandbox that lets Codex agents run directly inside PowerShell and other Windows developer tools, eliminating the need for WSL or virtual‑machine layers. According to the company’s developer portal, the sandbox enforces restricted token scopes, file‑system access rights, and dedicated sandbox user accounts, thereby containing each agent’s execution environment at the operating‑system level (OpenAI developers site). The sandbox code has been open‑sourced on GitHub, giving the community visibility into the security model and the ability to audit or extend the isolation mechanisms themselves.

The Windows version arrives only weeks after the macOS launch, which logged more than one million downloads in its first seven days—a figure OpenAI disclosed in its product announcement (OpenAI). That rapid uptake pushed Codex’s total weekly active user base to over 1.6 million, according to the same source. The demand was evident even before the Windows rollout: OpenAI reported that more than 500,000 developers had signed up for the Windows waiting list (OpenAI). By making the app available across all ChatGPT subscription tiers, OpenAI is positioning Codex as a universal productivity layer rather than a premium add‑on.

Codex’s core functionality hinges on asynchronous multi‑agent execution and “Automations” that delegate repeatable tasks across a project. Agents can be linked to external tools and workflows via “Skills,” allowing developers to maintain context while stepping into or guiding an agent’s work (OpenAI). The Windows sandbox preserves this workflow continuity by granting agents direct access to native Windows shells, which reduces latency compared to routing commands through a Linux compatibility layer. This design choice also aligns with enterprise security policies that often restrict cross‑OS interactions.

Open‑source publication of the sandbox has already sparked community contributions. Early pull requests on the GitHub repository focus on extending file‑system permission granularity and integrating additional Windows shells such as Windows Terminal. While OpenAI has not yet disclosed performance benchmarks, the native integration is expected to lower overhead for file‑intensive operations—a common bottleneck in code‑generation pipelines. The move also signals OpenAI’s broader strategy to embed its AI tooling deeper into developers’ existing environments rather than relying on browser‑only interfaces.

VentureBeat’s coverage notes that the surge in Codex adoption could pressure OpenAI to introduce usage limits for free and “Go” plan users, although no formal policy changes have been announced (VentureBeat). The company’s decision to make the app available to all ChatGPT plans suggests a willingness to monetize through higher‑tier subscriptions while keeping the entry barrier low. As competitors like Anthropic roll out agent‑centric features with larger context windows, OpenAI’s emphasis on OS‑level sandboxing may become a differentiator for developers who prioritize security and seamless integration over raw model size.