Use ox from Claude and ChatGPT with its MCP server

Connect ox to Claude or ChatGPT and ask in plain words: "deploy my repo", "why did that deploy fail", "put it on my domain". The assistant uses ox's tools, a remote MCP server at https://deploywithox.com/mcp. You stay in charge: every step that needs your browser comes back as a link for you to open, and no tool reads a variable you saved.

You ask your assistant in plain words, and it calls ox's tools at deploywithox.com/mcp. The steps that need you come back as links you open, and a secret is typed in the dashboard, never in the chat. The agent on your server builds the app and runs it as a systemd service behind Caddy, which serves it over HTTPS.

What you need

An ox account. You get one when you sign in with GitHub.
Nothing else to start. If you have no server yet, or have not connected GitHub, the assistant gives you the link for each. The quickstart shows the same steps by hand.

Connect in Claude

This works in Claude on the web and in the desktop apps.

Open Customize, then Connectors, and press Add custom connector.
Enter https://deploywithox.com/mcp and press Add.
Claude opens ox's sign-in page. Sign in with GitHub, read what the assistant may do, and press Allow.

On a Team or Enterprise plan, an Owner adds the connector for the organization, under Organization settings, Connectors, Add, Custom. Each member then finds ox under Customize, Connectors with the Custom label and presses Connect. If you are a member, ask an Owner to add it.

Connect in ChatGPT

Go to ChatGPT Plugins.
Select the plus button, then Add custom MCP server.
Enter a name and a description, such as "ox" and "Deploy to my own server".
Under Connection, enter https://deploywithox.com/mcp as the MCP server URL.
Set up the sign-in, read the risk warning, and select I understand and want to continue. ChatGPT opens ox's sign-in page: sign in with GitHub, read what the assistant may do, and press Allow.
Select Create as a plugin.
Look over the tools ChatGPT found on the server, and install the plugin.

To use it, start a new chat, type @ and pick ox.

Your ChatGPT plan or workspace decides whether you may add a server of your own.

Connect in Claude Code

Add ox with one command:

Add ox to Claude Code

claude mcp add --transport http ox https://deploywithox.com/mcp

Then run /mcp in Claude Code, pick ox, and sign in. Your browser opens ox's sign-in page: sign in with GitHub and press Allow.

Two skills teach Claude how to use ox, and each is one file you can give it: /ox-deploy-skill.md is the path from a repository to a live address, and /ox-skill.md has the ox.toml rules.

With a token

A script, or a client that cannot open a browser, can send an ox token in a header. Make a token with only the access the assistant needs. A read token lets it look: projects, deploys, logs, variable names and domains. A deploy token also lets it deploy and roll back a project you already have. A full token lets it use every tool: making a project, answering the first-deploy review, and adding a server or a domain need this one. The token is shown once and works for 90 days.

Make a token

ox tokens create --name claude --access full

This one is full because the first thing to ask below is a first deploy, which makes a project. For an assistant that only redeploys, make a deploy token.

You can also make it in the dashboard, under Settings, Sign-in and tokens. The first line below reads the token without showing it, so it stays out of your shell history:

Add ox with a token

read -rs OX_TOKEN
claude mcp add --transport http ox https://deploywithox.com/mcp --header "Authorization: Bearer $OX_TOKEN"
unset OX_TOKEN

Claude Code keeps the token in its own settings on your computer.

What the assistant can and cannot do

It can:

say what to do next toward your first deploy, and give you the link for each step that is yours;
list your servers, repositories and projects, and read a project's state, address, deploys and domains;
make a project from a repository on one of your servers, answer the first-deploy review and start the deploy;
deploy again, wait for the result, and read the deploy's log and your app's logs;
add a domain and list the DNS records to create;
roll back to the previous release.

Two tools change what is live: deploy and rollback. ox marks both that way. Most assistants ask you before they run a tool with that mark. Whether yours asks is the assistant's choice, not ox's.

It cannot:

delete anything: a project, a server, a domain, data or a backup;
read the value of a variable you saved. It sees the names. The first-deploy review can show the example value of a key that is not a secret, and log text hides what the Logs tab hides;
take a secret from the chat. You type a secret in the dashboard, or ox generates it;
run a command on your server, look into a database, or download or restore a backup;
change your account, your plan, your tokens or your settings.

Three steps are always yours, in your browser, through a link the assistant gives you: adding a server, connecting GitHub, and typing a secret. Everything the assistant cannot do stays in the dashboard and the CLI.

Three things to ask

"Deploy github.com/me/my-app to my own server with ox." The assistant checks your account, sends you the links it needs you to open, makes the project, answers the review and waits until the app is live. Then it gives you the address.

"My last deploy on ox failed. What went wrong?" It reads the deploy's log, shows you the lines with the error and says what to fix. The release that was live keeps serving. Troubleshooting lists the common causes.

"Put my ox project shop on shop.example.com." It adds the domain and lists the DNS records for you to create. The HTTPS certificate arrives by itself once the name points at your server.

What the assistant sees

ox answers a tool with names, states, addresses and log text. Log text is written by your own app, so ox marks it as output and tells the assistant to treat it as data, never as instructions. ox gets the tool calls: which tool, what you asked it for, and the name of the assistant's app. Every change the assistant makes is in your account's audit log (ox audit). What your assistant's maker does with the conversation is under its own terms. The privacy policy has the details.

Disconnect

In ox, open Settings, Connected apps. Each assistant you allowed is listed there with a Disconnect button, and it is out at once.

Then remove ox from the assistant. In Claude, open Customize, Connectors, and press Remove on ox. In ChatGPT, remove it under Plugins. In Claude Code:

Remove ox from Claude Code

claude mcp remove ox

If you connected with a token, revoke it: ox tokens revoke claude, or Settings, Sign-in and tokens.

Get help

When a tool fails, its answer says what happened and what to do next. For a failed deploy, start with Troubleshooting. For how ox protects your server and what it can reach, read Security. To write to ox, the contact address is in section 17 of the terms.