Skip to main content
OpenAI Codex speaks MCP, so the Gaia connector works there the same way it does in Claude Code — two commands, then you’re building.

​
Add the connector

1

Register the server

2

Sign in

Your browser opens, you sign in, and you pick which account to connect. The token is stored in your OS keyring.
The consent screen has a checkbox — “Also let it write code in your apps.” Leave it off unless you want author mode; you can turn it on later from Account → Connections without reconnecting.
3

Check it's wired up

Then, in a session:

​
Build something

Drop a spreadsheet in an empty folder and ask:
Codex imports the workbook, walks the contract conversation with you, then runs the build pipeline. When the harness needs a decision it presents the options as a numbered list in chat — reply with the number or the option name and it continues.
No spreadsheet handy? Download the event planning sample and use that.
Then keep talking:

​
Long builds

build takes 90–180 seconds and a chat turn can take over a minute. If your client doesn’t send a progress token, the connector hands the wait back after about 50 seconds with outcome:"running" rather than risk the call being abandoned. That is not a failure. The right response is a status call, which Codex should make on its own.
Never ask Codex to run the build again to check on it. The first build is still running server-side, so a second start isn’t a retry — it creates duplicate work someone has to clear by hand. The same goes for re-sending a chat message.

​
Manage the connection

You can also review or revoke the grant from Account → Connections.

​
How Codex differs from Claude Code

Same connector, same tools, same harness — only the client’s interface differs:
The connector is an MCP server, not a shell command, so there are no per-command approval rules to configure. Codex asks once when a tool is first used, and that’s it.

​
Next steps

Tool reference

Every tool the connector exposes.

Claude Code

The same connector, in Claude Code.