No MCP? Paste one sentence into any coding agent that can read a web page and make web requests. It follows a page that explains every step and does the same work:
Install
The server runs withnpx, so there is nothing to install by hand. The first start downloads one package of about 250 kB.
- Claude Code
- Cursor
- Codex
- OpenCode
- Other clients
The plugin installs the tools and the skill together.Start a new session. The plugin asks for nothing.To add only the tools:
antideploy-mcp@0.2, which takes every 0.2.x fix. To move to a new minor version, change that number.
The skill
Theantideploy skill tells the agent when Antideploy fits, how to connect, how to deploy, what to read when something fails, and when to use a different host instead. Claude Code, Cursor, Codex and OpenCode all read the same SKILL.md format.
The Claude Code plugin installs it for you. For the others, copy the folder skills/antideploy from the repository to .agents/skills/ in your project or ~/.agents/skills/ for every project; Cursor, Codex and OpenCode all look there. Keep the folder name antideploy.
Your first deploy
1
Ask for it
In your agent, say what you want, for example Deploy this project on Antideploy.
2
Approve one link
The agent calls
login and shows you a link and a short code. Open the link, check that the code matches, and click Approve. If you have no account, the same link creates one.3
Let it finish
The agent creates the application for the folder, creates a Postgres database if the project needs one, deploys, and polls until the app is live. It tells you the address and which Antideploy account it went to.
~/.antideploy/config.json, readable only by you. You can revoke it at any time on the tokens page.
What the agent can do
Deploys are asynchronous.
deploy returns a task id, and the agent calls deployment_status until it reads succeeded or failed. Each call waits up to 20 seconds for the deploy to finish, so it does not need to poll in a loop; a first deploy usually takes one to three minutes. Warnings and hazards in that answer are structured data meant to be read out to you.
Deleting an application or its database, custom domains and billing are not offered to agents. You do those in the dashboard.
Files, secrets and keys
- Tokens and passwords stay out of the conversation. The tools that write
.envanswer with the variable names only. - A
.envfile is never uploaded. To store variables from it, the agent callsset_env_from_filewith the names it needs. It reads the file on your machine and sends only those. It refuses the names the platform sets itself, such asDATABASE_URL. - Keys only you have. The agent calls
request_keys, which gives you a page on antideploy.com to paste a Stripe key or a bot token into, so the agent never sees it. - A folder is tied to one application. A folder without
.antideploy.jsonis refused, never matched to an application by guesswork. - Only the project folder. The tools act inside the folder the agent was started in. A folder named in a call must be inside it, a home directory or a drive is refused as a project, and nothing is written through a symbolic link, so an instruction hidden in a repository cannot send another folder to Antideploy.
Limits
The tools read the upload limits from the platform when they deploy, so they follow the API contract: currently 4,000 files and 28 MB for a project, and 28 MB for any one file. If a file is over the limit the agent is told which one, before anything uploads. One deploy at a time per application. A larger project can deploy from GitHub.Update and remove
In Claude Code:antideploy entry (and the skill folder) to remove.
Removing the tools does not log you out. Delete ~/.antideploy/config.json, or ask the agent to log out. To revoke the login on the server, use the tokens page.
A project key, for a pipeline
For CI or a second machine, create a key for one application in the dashboard and setANTIDEPLOY_API_KEY in the server’s environment. The tools then work without login, bound to that application. Treat it like a password: it is built to survive being committed, but you should not publish it.
Compared to the other ways in
If something goes wrong
- The tools do not appear. Check that the server started:
claude mcp listin Claude Code,codex mcp listin Codex,opencode mcp listin OpenCode. The first start needs a moment to download the package, and the server needs outbound HTTPS to antideploy.com. - The agent says it cannot reach antideploy.com. Its sandbox blocks the domain. Allow it in the agent’s network settings, or run the agent on your own machine.
- A deploy failed. The agent reads the reason and the logs, fixes the project and deploys again. See Why did my deployment fail?.
API reference
The MCP server is a thin client over the HTTP API. Anything it can do, a script can do with
curl.