Monitoring for OpenAI Codex apps

    Your OpenAI Codex app is live. Who tells you when it isn't?

    Codex opens the pull request, the tests pass, you merge. Production is a different place from the sandbox the agent worked in, and it is where things break.

    Where it runs: you deploy what Codex produced: a cloud host, a container platform, a VPS.

    Paste your app's address. We propose the checks, you say yes:

    We read the page once and propose the checks it needs. Nothing is created until you say so.

    Or just check it once, no account:

    No account needed. One check, from Frankfurt, in about twenty seconds.

    What breaks after launch

    None of these show up in the editor. All of them show up to your users first.

    The sandbox is not production

    Codex works in an environment with the dependencies it installed. Production has the versions the last deploy left, the environment variables ops set, and the network the firewall allows.

    Merged and unwatched

    Ten pull requests a day means ten deploys a day. Each one is a chance for a route to start returning 500, and the merge notification is the last time anyone looks.

    The certificate, the domain, the DNS

    None of it is in the repository, so none of it is in a pull request, so none of it is ever reviewed until it fails.

    Outside services

    The agent calls whatever API the task needed. Each key, quota and outage is now yours.

    The job that stopped

    Anything scheduled fails without touching the page users see.

    The email problem

    Mail sent by an app Codex wrote goes out through a provider client the agent chose, with the key in an environment variable and the domain configured in the provider's dashboard. When the send fails in production, the code path the agent tested in its sandbox does not run at all: the variable is unset, the provider refuses the key, the queue is not running. The page keeps working. ServerInspector checks the app from outside on a schedule, tracks the domain the mail is sent from, and on Pro runs an API test against the sending endpoint so a broken send is an alert rather than a support ticket.

    How to set it up

    Nothing is installed in the app. Monitoring happens from outside, the way your users reach it.

    1. 1

      Type your app's address into the free check above. It tells you in twenty seconds whether the site answers, how fast, whether the certificate and domain are in order, and which security headers are missing.

    2. 2

      Create a free account and add the same address as a monitor. It is checked every five minutes (every minute on Pro), and you get an email when it stops answering and another when it is back.

    3. 3

      Add your custom domain under Domains. Expiry is tracked and you are reminded 30, 14, 7 and 1 day before renewal.

    4. 4

      Add a health endpoint as a monitor; have Codex write one that checks the database and the queue if there is none.

    5. 5

      Mark deploys from CI with a deploy token so a restart does not page you and a failed deploy does.

    Free plan: 10 monitors or domains, 5-minute checks, commercial use welcome, no card.

    Questions OpenAI Codex builders ask