How to Configure ChatGPT Codex config.toml in AI Skill Notebook

When you use ChatGPT Codex, it may ask for your approval before it does things like:

• sending a network request, or

• editing files on your computer.

If you want Codex to run fully automatically, or you want to allow it to use other tools on your system, you can change these behaviors in Codex’s configuration file.

Important: Settings are per workspace.
Each workspace can have its own config.toml with different rules.

Where to find the config.toml file

If you’re using AI Skill Notebook, the file is located here:

<your-workspace>/codex/config.toml

You can open it inside AI Skill Notebook:

Top-right → Settings (icon) → Workspace Settings → ChatGPT Codex Configuration → Open

A simple example

Create or edit your config.toml to include something like:

approval_policy = "never"
sandbox_mode = "workspace-write"

approval_policy (asks you for permission or not)

This controls whether Codex should ask for confirmation before it performs an action.

AI Skill Notebook uses never by default (no prompts), but you can choose:

• untrusted — only runs known-safe actions

• on-failure — runs in a safe mode by default; asks only when higher permissions are needed

• on-request — Codex decides when to ask

• never — never asks; runs automatically

sandbox_mode (how much access Codex has to your files)

This controls where Codex is allowed to read/write files.

AI Skill Notebook uses workspace-write by default, but you can choose:

• read-only — can read files, but cannot modify anything

• workspace-write — can only modify files inside the current workspace folder

• danger-full-access — can modify files anywhere on your computer (use with caution)

Using third‑party MCP tools (Model Context Protocol)

Some third‑party tools support MCP, which lets Codex control them through natural language.

Example: Playwright can automate a browser. Programmers often use Playwright with code, but Playwright MCP lets Codex operate the browser based on instructions like “open this site and log in”.

Enable Playwright MCP in config.toml

Add a new section under mcp_servers like this:

approval_policy = "never"
sandbox_mode = "workspace-write"

[mcp_servers.playwright]
command = "npx"
args = ["-y", "@playwright/mcp@latest", "--headless"]
timeout_ms = 60000
retries = 2

What this does:

• Starts the Playwright MCP server via npx

• Runs it in headless mode (no visible browser window)

• Sets a 60-second timeout per request, and retries twice if needed

More configuration options

For the full list of settings, see the OpenAI Developers documentation here:

https://developers.openai.com/codex/config-basic