Codex MCP: Two Commands, One Live AI Agent

Codex MCP takes two commands, and they are right below. Most guides then point Codex at GitHub, Sentry or a database. This one points it at your customers. If you build your own site with Codex, the same session can create the AI agent that answers them, wire it into your layout and write the API that agent calls. You review a diff, like anything else Codex does.

Codex MCP: how do I add a server?

You add a remote MCP server to OpenAI Codex with codex mcp add and its URL, then start the sign-in with codex mcp login. A server with OAuth needs no key: a browser tab opens and you approve there. For your Ultimo Bots agent:

codex mcp add ultimo-bots --url https://mcp.ultimo-bots.com/mcp
codex mcp login ultimo-bots

In the browser tab you sign in to Ultimo Bots, choose one workspace and the permissions, and click Allow connection. For Codex to create an agent, keep All agents in this workspace, including future agents and the permission Create and rename agents. Then restart Codex, because the tools load with the next session. codex mcp list shows your servers, and /mcp shows them inside the terminal interface.

How do I use MCP in the Codex macOS app?

In the desktop app, which OpenAI now calls the ChatGPT desktop app, open Settings, then MCP servers, select Add server, choose Streamable HTTP, enter the URL, save and restart. Select Authenticate when it asks for the sign-in. If you already ran the commands, you can skip this: OpenAI's Codex documentation says the ChatGPT desktop app, the Codex CLI and the IDE extension share this configuration (read November 8, 2026).

What MCP servers work with Codex?

Codex works with local STDIO servers it starts itself and with remote servers over Streamable HTTP. Ultimo Bots is a remote one, so there is nothing to install or keep running on your machine. No browser on that machine? codex mcp login ultimo-bots --no-browser prints a URL to open elsewhere.

Our server has more than 100 tools, and two lines in config.toml keep Codex to the ones you want. Once the agent is set up, enabled_tools is an allow list, and default_tools_approval_mode = "writes" makes Codex prompt you for every tool that is not marked read-only. Our read tools carry that mark, so this version reads freely and can never change a setting:

[mcp_servers.ultimo-bots]
url = "https://mcp.ultimo-bots.com/mcp"
default_tools_approval_mode = "writes"
enabled_tools = ["list_agents", "list_conversations", "get_conversation", "list_leads", "get_lead", "get_insights"]

Every other option is in our Codex guide in the help center, and the portal's MCP connections page has the commands ready to copy, with Revoke access next to each connection. Plugins in ChatGPT chats are a separate setup, so ChatGPT MCP has its own steps.

One session: from one prompt to a live agent

Ultimo Bots is an AI agent platform you can drive from a terminal and extend in plain words: its MCP server handles the setup, and Mo, the builder in the AI Studio, turns a sentence into a tested skill on your API or a remote MCP server. Take Roast Lane, a fictional coffee subscription shop. Sam, the founder, built the site in Next.js with Codex and answers every customer question by email. Sam opens the project and writes one prompt: "Create an Ultimo Bots agent called Roast Lane. Add the shop, shipping and FAQ pages as knowledge and wait until they are ready. Set the widget colors from globals.css, then fetch the embed code and add it to the site-wide layout."

Each website page goes in as its own source, and Codex can list a site's pages first with discover_knowledge_links. Imports run in the background, so Codex polls get_operation instead of assuming they worked. get_embed_instructions returns the exact code for this agent, and Codex places it the way our embed guide describes for Next.js. You read the diff and deploy as usual.

A new agent starts on the Free plan with a small one-time allowance of free AI credits. It can go on your site right away, and when the credits are used up it pauses until you choose a plan in the portal.

What is the real benefit of MCP servers for Codex?

The real benefit of MCP is that Codex works from live facts outside the repo instead of whatever you paste in. A week after launch Sam asks: "Read this week's conversations and list what our site does not answer." Codex reads them with list_conversations and get_conversation, opens the matching page in the repo and proposes the missing paragraph. The fix ships with the next deploy, and refresh_knowledge_sources brings it into the agent. That loop runs on no AI credits, because Codex does the thinking on your plan with OpenAI.

Codex writes the endpoint, Mo gives the agent the job

Codex builds your side of the API, and Mo teaches the agent to use it. You could write the integration yourself, and then you would also own the prompt, the confirmation step and the tests, on every channel. Mo does that from a description in the AI Studio, so your time goes to your product. Sam's top question is whether a roast is in stock, and the answer changes daily. So Sam has Codex add two routes to the project: a stock list, and a waitlist that writes to the shop's own database, behind a key.

1. 1

   Mo reads the API

   It reads your public docs if you have them, or asks what the endpoints return. A short description from Codex is enough.

2. 2

   You paste the key

   Mo asks for the API key in a secure field, and it is stored encrypted.

3. 3

   Mo builds and tests

   It tries each call, the POST as a dry run, then runs a short test chat against your API and tells you what the agent can now do. That chat makes real calls, so delete any test row it leaves.

At 3:12 am in Sam's time zone it is lunchtime in Melbourne, and a customer asks about the Ethiopian roast. The agent checks the API: "That one is sold out, and the next roast is Tuesday. Want me to put you on the waitlist?" She gives her email, the agent reads it back, and after her yes it posts to /v1/waitlist. Sam wakes up to a new row in the shop's database and no email to answer.

The pattern fits any public HTTPS API that takes an API key or bearer token, and any remote MCP server that does. Codex can write either end of it. The AI Studio page lists the building blocks, and app actions lists the ready-made ones for Stripe, Cal.com, HubSpot and others.

AI credits pay for the agent's answers and for Mo, from one balance per agent: Smart $29 a month ($19 billed yearly) with 500 AI credits and 3 skills, Boost $59 ($39) with 2,000 and 10 skills, Ultimo $249 ($199) with 10,000 and 30 skills. Teammates are free (pricing).

One branch, ten minutes

Open a branch, run the two commands, paste Sam's prompt with your own page names, and read the diff. Nothing reaches your visitors until you merge it, and ten minutes tells you whether you like the result. When the agent is live, give it a few days of real questions and pick the one that needs live data from your system. That question is your first message to Mo. Using other clients too? See Claude MCP and Cursor MCP.

Questions and answers