Feature request: Let ChatGPT Web connect to local Codex with CLI-style permission profiles

Feature request

I would like ChatGPT Web to be able to directly control my local computer through Codex, using the same kind of permission profiles available in Codex CLI.

Right now, Codex CLI can run locally and operate on my machine with configurable permissions, including broad/full-access modes when the user explicitly chooses them. But ChatGPT Web cannot use that same local-control pathway. This creates an unnecessary split:

• Codex CLI has local file/terminal/computer access.
• ChatGPT Web has the better conversational interface, context, memory, and long-running planning experience.
• Power users have to manually copy instructions between ChatGPT Web and Codex CLI instead of letting ChatGPT Web delegate execution to a local Codex agent.

Desired behavior

Add an opt-in mode where ChatGPT Web can connect to a local Codex app/daemon and execute tasks on my machine using explicit permission profiles, for example:

• read-only
• current workspace only
• selected folders
• terminal access
• GUI/computer-use access
• full computer control / danger-full-access equivalent

This should not be enabled by default. It should require local installation, explicit user login, local confirmation, and clear warnings.

Why this matters

For advanced users, ChatGPT Web is the natural command center. It is much easier to plan complex workflows, manage context, explain goals, and supervise agents from the web UI. But the actual execution often needs local machine access: editing files, running scripts, moving folders, opening apps, using the terminal, or operating the desktop.

The current separation makes workflows more complex than necessary. If Codex CLI already supports local execution with user-selected permissions, ChatGPT Web should be able to use the same local execution backend instead of forcing users to switch interfaces.

Proposed architecture

ChatGPT Web should act as the planner/orchestrator.

A local Codex app/daemon should act as the executor.

The local Codex executor should expose permission profiles and require the user to approve the connection/session locally. ChatGPT Web should never silently gain full OS control.

Useful safeguards:

• per-session permission selection
• visible local connection indicator
• full action logs
• emergency stop button
• confirmation for high-risk actions such as deletion, sending emails, installing software, accessing credentials, browser cookies, or system folders
• ability to restrict access to specific folders/apps
• ability to revoke the connection immediately

Expected outcome

A power user should be able to open ChatGPT Web, connect it to local Codex, choose a permission profile, and let ChatGPT directly operate the local machine through Codex, similar to what Codex CLI can already do locally.

This would make ChatGPT Web + Codex much more useful as a real agentic workstation while still preserving explicit user control and safety boundaries.