ChatGPT Work desktop adds cloud vs local run controls
OpenAI staff said ChatGPT Work runs in the cloud on web and mobile, while desktop can now choose cloud or computer execution. The clarification followed confusion about closed-laptop and local-environment behavior.

TL;DR
- ChatGPT Work now has an execution-locality split: web and mobile run in the cloud, while desktop Work can run in the cloud or on the computer, per reach_vb's clarification and dkundel's run picker.
- Closed-laptop agents are the hook: gdb's post says Work can keep running from mobile while a laptop is closed.
- The naming is still muddy: dkundel on the harness says Work uses the Codex harness in the background, while dkundel on environment config says ChatGPT Work does not expose environment configuration.
- OpenAI is still collecting field reports: thsottiaux asked for 50 to 100 Work users for early feature access after asking what people wanted from the app thsottiaux's cohort call thsottiaux's feedback prompt.
- Desktop agents are already pushing users toward isolation layers: kevinkern tested OpenAI Tart for Codex macOS VMs, while emollick showed Codex installing Blender on Windows with one permission click kevinkern's Tart test emollick's Blender test.
OpenAI's July 16 release notes describe a desktop patch that added Unified Recents, Projects, and cross-device continuation for cloud Work. The current ChatGPT Work and Codex Help Center page spells out the missing rule: cloud Work syncs across surfaces, local Work stays on the computer. The Codex pricing page has the billing gotcha: Work mode and Codex share usage, and local messages and cloud chats share a five-hour window. An OpenAI community thread shows why the patch landed as cleanup: users were trying to understand where Projects, chat history, Codex state, and ~/.codex fit in the new app.
Cloud and computer runs
The clean rule is now locality. OpenAI's Help Center page says Work on web and mobile runs in the cloud, while Work in the desktop app can also use local files and desktop apps with permission when the plan and workspace allow it.
The same page gives the state model:
- Cloud Work chats sync across web, mobile, and desktop.
- Work chats started on web or mobile appear in the desktop app.
- Cloud Work chats started on desktop can be continued on web or mobile.
- Local chats stay on the computer.
That matches reach_vb's clarification, which said Work runs in the cloud on web and app, while desktop Work runs locally. A later desktop change added the missing picker.
Desktop run picker
The desktop control is a Run this chat menu with two options: On my computer and In the cloud. dkundel said the choice had been available on desktop since Thursday in his run-picker reply.
The boundary is still product-specific. dkundel on the harness said ChatGPT Work can run locally or in the cloud, but that choice is unavailable in the Codex part of the app, even though Work uses the Codex harness in the background.
There is also a capability boundary inside cloud Work. dkundel on environment config said ChatGPT Work has no environment configuration and is useful for non-coding tasks that do not rely on a local filesystem.
Work, Chat, and Codex boundaries
The updated desktop app has a global switcher for ChatGPT and Codex, then a Chat and Work toggle inside ChatGPT. OpenAI's release notes say the July 16 desktop update made that layout live on macOS and Windows for all plans.
OpenAI's Work and Codex doc splits the surfaces this way:
- Chat: questions, search, brainstorming, quick conversational help.
- Work: research, analysis, documents, spreadsheets, presentations, reports, and Sites.
- Codex: code writing, debugging, tests, commands, repositories, and developer tools.
Codex remains a separate view with separate history. While inside Codex, dkundel's Quick Chat reply described Quick Chat as a fast ChatGPT surface without leaving Codex.
Work also inherits Codex economics. The Codex pricing page says Work mode and Codex share pricing, credits, and usage limits, and the FAQ says model choice, context, reasoning, tool use, retrieval, and caching affect usage.
Managed browser, plugins, and Sites
Work is being positioned around finished tasks, not just coding. thsottiaux listed four examples in his Work list:
- Creating and hosting Sites.
- Managing email.
- Summarizing large document sets.
- Creating docs, sheets, and slides.
The mobile example from reach_vb's browser post used a managed browser plus connected email and calendar context to find nearby restaurants and reservation details. reach_vb on plugins added that email and calendar are connected through plugins, and those connections are universal across surfaces once connected.
Sites is part of the same push. OpenAI's July 9 release note says Sites can be created from ChatGPT Work on web, or from Work or Codex in the desktop app, with public beta availability varying by plan and region.
Feedback queue
OpenAI is treating Work as an active feedback loop. thsottiaux asked what people were building in his weekend prompt, asked how ChatGPT could improve in his feedback prompt, then opened a 50 to 100 person early-access group for Work users in his cohort call.
The workflow example was pure delegation. In thsottiaux's dictated DM task, he asked Work to scan thousands of DMs, extract users who mentioned ChatGPT Work, build a spreadsheet, classify each workflow into 8 to 12 labels, and rate sophistication for cohort selection.
The complaints were concrete:
- ChatGPT and Codex connections felt brittle, with threads sometimes loading and sometimes failing, according to MatthewBerman's reply.
- Remote Codex in the ChatGPT app could not create new folders or projects in a preferred directory, chrisbbh's request said.
- Documentation was still hard to find, and simonw's documentation complaint said desktop Work looked like a Codex skin while mobile Work behaved like a separate mode.
VM isolation for desktop agents
Local execution brings local UI problems. zeeg called the Codex desktop app mostly nonfunctional on Windows plus WSL in his Windows note, while kevinkern tested OpenAI Tart so Codex could spawn an isolated macOS VM instead of stealing focus on the main desktop.
That isolation had a cost. kevinkern's Tart test put the VM at around 8 GB of RAM and slower control.
steipete described the other side of the tradeoff: Codex used browser and computer control to upload an image through GitHub's UI, and his PR upload post said he runs Codex in VMs so agents do not steal app focus.
Computer use is already broad enough to leave the coding path. emollick's Blender test showed Codex downloading and installing Blender on Windows, then making and animating a 3D otter after one Windows permission click.