Deploy from Codex to a Live URL with MCP

Yes, once it has a tool that does the deploying. Codex writes and runs code in your project but does not host the result. Add ProductEcho to Codex as an MCP server and one prompt is enough: the agent reads your project, creates a PostgreSQL database if the app needs one, uploads the code, deploys it, and gives you a public HTTPS URL.

It works the same in the Codex CLI, the desktop app and the IDE extension, because they share one MCP configuration. This page is about deploying apps and websites you built with Codex, not about installing Codex itself.

Codex keeps its MCP servers in ~/.codex/config.toml. From a terminal, one command writes the entry and a second signs you in:

codex mcp add productecho --url https://api.productecho.com/mcp
codex mcp login productecho

codex mcp add registers the server and may not sign you in by itself, so run codex mcp login productecho as well. It opens a browser window where you sign in and approve access to your ProductEcho workspace. The consent covers workspace discovery, deployments and database management, and no API key is involved. Run codex mcp list to confirm the entry, or type /mcp in a Codex session.

In the desktop app or the IDE extension

Or edit config.toml yourself

[mcp_servers.productecho]
url = "https://api.productecho.com/mcp"

A project can carry its own .codex/config.toml, but Codex only reads it for projects you have marked as trusted. After saving, sign in with codex mcp login productecho.

The MCP server gives Codex the tools. The skills give the agent ProductEcho's instructions for using them: how to inspect a project, what each runtime needs, and how to create and wire a database. Install them once:

npx -y skills add productecho/skills --skill '*' --yes --global --agent codex

This copies four skills into ~/.agents/skills, a folder Codex scans for your own skills. Codex picks up new skills on its own, and a restart helps if they do not show. Type $productecho-deploy to run one by name, or run /skills to see the list. You can also just ask in plain language, and Codex chooses the skill when the task matches.

Open your project in Codex and run:

Deploy this app to ProductEcho with PostgreSQL and give me the live HTTPS URL.

Drop "with PostgreSQL" if the app has no database. To call the skill by name, start the message with $productecho-deploy. When the deploy finishes, the agent prints the URL.

Server apps run as containers and single-page apps that build to static files are served from a CDN. The project inspection decides which one fits, and projects with a Dockerfile always run as containers. What each runtime needs:

• Node.js, Next.js and Express: runtime packages belong in dependencies, and the app must listen on PORT and bind to 0.0.0.0.
• Python (FastAPI, Django, Flask): a production server in requirements.txt and a start command, from a Procfile or one you confirm.
• Go: go.mod and go.sum at the project root, and the app reads PORT.
• Java (Spring Boot, Quarkus, Micronaut): the ./mvnw or ./gradlew wrapper in the project.

In a monorepo, the agent asks which project to deploy; see deploying one app from a monorepo.

ProductEcho flags deploys, environment updates, pauses and deletes as destructive. Codex's documentation says an MCP tool call with a destructive annotation always needs your approval, unless the tool is also marked read-only, so you see each one before it runs. Three habits keep that review meaningful:

• Read the prompt for deploy_application, update_application_env and any tool whose name starts with delete_. The sign-in grants write access, and that includes deleting apps and databases.
• Set default_tools_approval_mode = "writes" on the ProductEcho entry. Codex then prompts for every tool that is not marked read-only, which includes every deploy, environment, pause and delete tool.
• Take the delete tools away. disabled_tools is a deny list, so Codex will not call the tools named on it. delete_application and delete_postgres are the two that remove an app or a database.

[mcp_servers.productecho]
url = "https://api.productecho.com/mcp"
default_tools_approval_mode = "writes"
disabled_tools = [
  "delete_application",
  "delete_postgres",
]

Keep secrets out of the code. The upload skips .env files, so set values as environment variables and let the agent read them back with sensitive values masked. To go back to an earlier version, ask the agent to redeploy that version of your code.

ProductEcho does not show up in Codex

Run codex mcp list, or type /mcp in a Codex session. If ProductEcho is missing, check that ~/.codex/config.toml has a [mcp_servers.productecho] table with the url, then restart Codex. A server defined in a project's .codex/config.toml is ignored until you trust that project. A typo in a setting name or value stops Codex from loading the file, and its error message names the line.

The sign-in does not finish, or the tools say unauthorized

The server answers 401 until you approve access in the browser window. Run codex mcp login productecho and finish that step. In the desktop app or the IDE extension, select Authenticate on the ProductEcho entry. If it stays stuck, run codex mcp remove productecho, add it again and sign in again.

Codex runs on a remote machine and the browser cannot reach it

After you approve access, the browser is sent back to a local address that Codex listens on, http://127.0.0.1/callback by default. On a remote machine or a devbox your browser cannot reach that address. Set mcp_oauth_callback_url in ~/.codex/config.toml to an address your browser can reach and sign in again. ProductEcho accepts HTTPS and local callback addresses.

The skills are missing

Check that ~/.agents/skills/productecho-deploy exists, run the install command again if it does not, then restart Codex and run /skills.

The deploy finished but the URL shows an error

Ready means the container started, not that your routes work. A gateway error usually means the app listens on the wrong port: it must bind to 0.0.0.0 and read the PORT variable. Ask the agent to check the port and redeploy, or follow why a deployment returns 502.

A Python deploy was refused

Without a Procfile, a Python app needs a start command. The agent proposes one from your code and asks you to confirm it before deploying.