Expose Existing Per-Tool MCP Configuration in the Codex UI

Start by reading the existing per-tool permissions model and config.toml handling, then inspect the Codex Desktop MCP server settings entry point. Define the UI scope for exposing Automatic, Ask First, and Disabled per-tool choices, along with the requested discovery features. Done means the existing permissions can be managed graphically without changing the underlying model.

What version of the Codex App are you using (From “About Codex” dialog)?

Codex Desktop 0.1.44.5

What subscription do you have?

ChatGPT Plus

What platform is your computer?

Linux 6.1.0-51-amd64 x86_64 unknown

What issue are you seeing?

This is a feature request rather than a bug.

According to OpenAI Support, Codex already supports per-tool permissions through its existing permissions model (including config.toml).

However, these capabilities are not exposed through the graphical interface.

When working with MCP servers exposing dozens or hundreds of tools, configuration becomes difficult to discover and manage because it requires editing configuration files.

I would like to see these existing capabilities exposed directly in the Codex UI.

Suggested improvements:

• Per-tool permissions (Automatic / Ask First / Disabled)
• Permission presets (Read, Search, Write, Dangerous)
• Tool search/filter
• Remember approval decisions
• Explain why a tool is being requested
• Lightweight tool activity history
• Basic per-tool metrics (calls, success rate, average duration)

This request is about improving usability and discoverability, not changing the existing permissions model.
Thank you.

What steps can reproduce the bug?

1. Open Codex Desktop.
2. Configure an MCP server exposing multiple tools.
3. Open the MCP server settings.
4. Try to configure permissions for individual tools.

Result:
There is currently no graphical interface to manage per-tool permissions. Users must rely on configuration files (config.toml), even though the underlying permission model already supports per-tool configuration.

What is the expected behavior?

Codex should expose the existing per-tool permission model through the graphical interface.

For each tool, users should be able to select:

• Automatic
• Ask First
• Disabled

Additional usability improvements that would be valuable:

• Permission presets (Read, Search, Write, Dangerous)
• Tool search/filter
• Remember approval decisions
• Explain why a tool is requested
• Lightweight tool activity history
• Basic per-tool metrics

This request is about improving discoverability and usability of existing functionality rather than introducing a new permission model.

Additional information

OpenAI Support confirmed that Codex already supports per-tool permissions through its existing permissions model (including config.toml).

This request is specifically about exposing those existing capabilities through the Codex UI.

I am submitting it using the bug template because I could not find a dedicated feature request template.

Thank you for considering this suggestion.

João Alves
Portugal