For AI agents: the index of this site is https://dejo.run/llms.txt, and every page is also served as markdown (add .md to its address, or send Accept: text/markdown).

dejo

Codex browser: what works in the CLI and IDE

BlogBy Rohit Kashyap

I build Dejo, one of the browser servers below, so weigh the last part with that in mind. The facts about Codex come from OpenAI's own docs, read on 13 November 2026.

Short answer: OpenAI's built-in Browser works in ChatGPT on the web and in the ChatGPT desktop app, not in the Codex CLI or the IDE extension. In OpenAI's words: "Browser isn't available in Codex CLI or the Codex IDE extension. Open the ChatGPT desktop app to use the built-in browser." To give Codex a browser in the terminal or your editor, add a browser MCP server to ~/.codex/config.toml.

Where Codex can use a browser

From OpenAI's Browser docs:

where you use Codexbrowser
ChatGPT on the web, ChatGPT desktop appthe built-in Browser: opens websites, gathers current information, acts while you watch
Codex Cloud tasksa browser inside the cloud task, to inspect your web app, reproduce a bug or check a fix
Codex CLI, Codex IDE extensionnone built in; add a browser MCP server

The same docs give one safety rule worth keeping whichever browser you use: treat page content as untrusted, and review the site and the proposed action before sharing anything sensitive.

Add a browser to Codex CLI in a minute

Codex reads MCP servers from ~/.codex/config.toml. Either run codex mcp add, or write the table yourself.

Dejo (remembers each task, Python 3.11+):

pip install dejo-mcp
playwright install chromium
codex mcp add dejo -- dejo-mcp

or in ~/.codex/config.toml:

[mcp_servers.dejo]
command = "dejo-mcp"

Playwright MCP (Microsoft's, Node):

[mcp_servers.playwright]
command = "npx"
args = ["@playwright/mcp@latest"]

Check with codex mcp list, and inside Codex with /mcp. Codex waits 10 seconds for a server to start unless you set startup_timeout_sec. The MCP config generator writes the block for any server.

Which one to add

We measured the repeat cost in Claude Code rather than Codex: on 19 real-website tasks, repeat runs with Dejo averaged 89K tokens against 409K per run with Playwright MCP. The method is on the benchmark page; we haven't run the same test in Codex yet.

Signing in

Sites behind a login need a session the browser keeps. With Dejo, Codex calls dejo_login, a window opens, you sign in yourself, and the session stays in Dejo's profile; details in keep your AI agent signed in. With Playwright MCP, see Playwright MCP login.

The full Codex setup page, with timeouts and settings: browser automation in Codex.

Related reading: Playwright MCP login, why Playwright MCP uses so many tokens and Dejo vs agent-browser.

FAQ

Does Codex have a browser?

Not in the Codex CLI or the Codex IDE extension. OpenAI's docs say: "Browser isn't available in Codex CLI or the Codex IDE extension. Open the ChatGPT desktop app to use the built-in browser." Codex Cloud tasks can use a browser to inspect a web app. In the CLI and IDE you add a browser as an MCP server.

How do I give Codex CLI a browser?

Add a browser MCP server. For Dejo: pip install dejo-mcp, playwright install chromium, then codex mcp add dejo -- dejo-mcp, or a [mcp_servers.dejo] table in ~/.codex/config.toml with command = "dejo-mcp". Playwright MCP works the same way with npx @playwright/mcp@latest.

Which browser MCP is best for Codex?

For one-off pages, Playwright MCP is fine. For website tasks Codex will repeat, a server that remembers the task, such as Dejo, makes the next run one tool call instead of a full exploration. In our Claude Code benchmark, repeat runs with Dejo averaged 89K tokens against 409K with Playwright MCP.