OpenAI Codex

Codex gives you the most reach for a single setup: the Codex command-line tool, the Codex IDE extension and the ChatGPT desktop app all read the same configuration, so provisioning a device once puts your connectors in front of all three.

How it works

• Skills land in Codex's personal skills folder.
• Connectors are added through Codex's own codex mcp command rather than by editing its configuration file directly. That file holds your own Codex settings and backs all three surfaces, so Harriet lets Codex write it and never touches it itself.

Set it up

1. Enable the target. Turn on OpenAI Codex in Settings.
2. Install the Harriet Endpoint agent on each machine and sign in.
3. Check your connectors with codex mcp list, and in the ChatGPT desktop app.

If neither the codex command nor the ChatGPT desktop app is installed on a machine, Harriet skips the connector step there and still deploys your skills — they appear as soon as either is installed. The ChatGPT app ships its own codex binary; Harriet uses that when there is no separate CLI on PATH.

Removing a device

Removal takes Harriet's connectors back out the same way. If Codex has already been uninstalled by then, Harriet cannot remove them and instead records the exact codex mcp remove commands to finish by hand. The entries are inert either way, because the credential they carry is revoked. See removing a device.

Using Harriet as Codex's model provider

This is optional and manual. Codex reads its model provider from its configuration file, which Harriet deliberately does not edit. Copy the base URL and key from either the device in Endpoint AI (Connect Codex to the Harriet gateway) or Harriet Endpoint on the machine (under Agents, Codex's ⋮ menu → LLM gateway URL), then paste them into that file yourself.

Uncheck OpenAI Codex under Agents in Harriet Endpoint if you do not want Harriet to apply skills and connectors on this machine.

Read next