OpenAI adds voice control and multi-folder projects to Codex

OpenAI's July 23 Codex changelog turns voice into a live control layer for ChatGPT desktop work. The update says ChatGPT Voice, powered by GPT-Live, can start, check, and steer tasks across Chat, Work, and Codex in the desktop app, while local projects can now include multiple related folders.

That combination matters more than a routine UI feature. It gives paid desktop users a way to coordinate agent work verbally while keeping a project spread across a primary repo, shared packages, documentation, and adjacent services.

Codex gets a voice loop for active work

OpenAI says the new voice mode is available in the ChatGPT desktop app for Plus, Pro, Business, Edu, and Enterprise plans, and through Remote on iOS. A user can start a new chat or task in voice mode, then ask ChatGPT to start, inspect, or redirect work in other threads.

On macOS, the feature can also use Screen context to share an appshot of the frontmost window. That makes the voice loop more useful for real development sessions: the assistant can respond to what is on screen, not only to a detached spoken instruction.

TestingCatalog reported related bundle signals on July 23, describing a broader realtime voice layer for Codex-style task control. Its coverage should be treated as early product telemetry, not as a replacement for the official changelog. The official OpenAI entry is the stronger source for what is available now.

Multi-folder projects reduce repo friction

The second part of the release is quieter, but practical. Local projects in the ChatGPT desktop app can now include multiple related folders. Users add folders from the project menu, then choose the primary folder.

OpenAI says new chats, Git operations, and automatic discovery of AGENTS.md, skills, and config.toml use that primary folder. Secondary folders remain available for file search, reading, and editing. For teams working across a frontend, backend, shared package, and deployment folder, that is a cleaner model than repeatedly attaching context or switching projects.

The primary-folder rule is also the caveat. Git operations and automatic config discovery do not become equally active everywhere. Users should pick the folder that owns the branch, instructions, and command surface they expect Codex to treat as the center of the task.

Who can use it now

The release is not a universal ChatGPT feature. OpenAI lists voice availability for Plus, Pro, Business, Edu, and Enterprise plans in the desktop app, plus Remote on iOS. It is a Codex and ChatGPT desktop workflow update, not a new API model launch.

The multi-folder project support applies to local projects in the ChatGPT desktop app. The changelog does not say that cloud environments, API workflows, or every connected workspace mode inherit the same behavior on day one.

For developers, the immediate test is simple: create a local project with the main repository as the primary folder, add adjacent folders that Codex needs to read or edit, then use voice for steering rather than for detailed patch instructions. Voice is best for intent, triage, and course correction; precise file changes still need visible review.

What to watch next

This release sits beside a wider shift toward agent operators that run across chat, files, screens, and remote devices. The useful question is whether OpenAI keeps voice as a steering layer, or expands it into a fuller hands-free operator for desktop tasks.

Teams evaluating agent workflows should pair this update with their normal review rules: keep repository instructions explicit, limit sensitive folders, and check generated changes before merge. LinkLoot's AI agent tools guide is a good companion when comparing Codex against browser, IDE, and cloud-agent alternatives.

For now, the concrete change is clear: Codex can be steered by voice in ChatGPT desktop, and local projects can span multiple folders without losing a primary workspace anchor.

Sources and methodology

This post uses OpenAI's Codex changelog as the primary source for availability and behavior. TestingCatalog is used as independent early coverage for the realtime voice direction, with its claims kept separate from the official release details.