Skip to content
Adrythm

Production checklist

Before real customer data goes into a Replit app

Replit Agent can get an app working quickly. These are the things to settle before customers, staff or payments depend on it.

Why read this and not just the docs

Replit documents each of these on its own page, written to explain the feature. This is one checklist, written to protect the business: it puts them in the order they matter once real data is involved, adds the two decisions no documentation can make for you, who owns the workspace and who reviews the code, and adds what went wrong in the five web apps Adrythm runs on Replit.

6 steps, in order.

  1. 01

    Keys go in Secrets, never in the code

    Replit's Secrets tool encrypts API keys, tokens and database connection strings and hands them to the app as environment variables. Check who can see them: Replit's table shows collaborators and organization members with the Owner role seeing secret values, other organization members seeing only names, and people who remix someone else's project seeing names at most.

    • Replit's Secrets page says secrets reach every deployment type except Static. An app that needs a key at run time cannot be published as a Static site.
  2. 02

    Use a real database, and know how far back it can go

    Every Replit App works with two databases: a development database you and Agent build with, and a production database that serves the published app. Agent can change the development one but not production, and schema changes reach production when you publish. Replit's data recovery page, read 2026-11-11, gives the windows: development rolls back to any Agent checkpoint up to 7 days old; production has point in time recovery for 7 days, up to 28 on Pro, Teams and Enterprise; daily scheduled backups of production exist but are off by default.

    • Turn on scheduled backups, and keep a copy of the data outside Replit as well. A deleted record noticed on day 8 is outside the default window.
    • Try a restore once, on development, before you need one on production.
    • Publish the code and the schema change it needs together, and check production straight after. On 2026-10-21 a publish of one of the five apps Adrythm runs on Replit put live a feature that read a table the production database did not have yet. The feature failed on every call until the table was applied.
  3. 03

    Choose how the app is published, and what proves it is up

    Replit documents four publishing types: Autoscale, which adjusts to usage; Static, for sites that do not change with user input; Reserved VM, which runs continuously on a fixed amount of computing power; and Scheduled, which runs at times you choose. A website and an always on automation server usually want different ones. All five web apps Adrythm runs on Replit publish as Autoscale, read from their configuration files on 2026-11-11.

    • Make the health check run one database query. On 2026-10-28 the database connection pool in one of those apps stalled: every request that needed the database hung, while the health address, which only checked that the server was up, kept answering 200. Replit's database page offers a Monitoring tab, for production databases only, that shows the queries running at that moment.
    • Read Replit's own run notices with care. In another of the five, every publish left the production server holding the port the development server wanted, so the development run failed and the editor reported the app as not running while the live site served normally. The fix was a run script that checks which process owns the port.
  4. 04

    Decide who owns the workspace, the bill and the code

    Settle whose Replit account the app lives in, who pays for it, and who can add or remove people, before the app holds anything important. Moving ownership later is possible, but it is easier when nobody is depending on the app yet.

    • Check that the code installs somewhere other than Replit before you count on being able to leave. In one of the five apps, read 2026-11-11, the package-lock.json written on Replit pointed 152 of its 207 package downloads at Replit's internal package firewall, http://package-firewall.replit.local, which nothing outside Replit can reach. A plain npm ci elsewhere fails until those addresses are rewritten to the public npm registry. The four apps that use pnpm had no such addresses in their lockfiles.
  5. 05

    Put a ceiling on the agent bill

    Replit documents usage limits that block usage based services once they are reached, a Plan Mode where Agent plans without changing code, and Agent modes that trade cost against capability. Set a limit before a long build, not after the invoice.

    • Teams that already pay for their own AI coding tools can connect an editor to the app over SSH, which Replit offers on its paid plans.
  6. 06

    Logins, roles and a second pair of eyes

    A first version often has one user who can do everything. Before launch, decide who can see and change each kind of record, then have somebody who reads code check the parts that handle money, personal details and permissions.

    • Decide where review happens. In the apps Adrythm runs, Replit Agent commits straight to the branch the app runs from, so its changes are reviewed after the fact. Other changes go through a branch and a pull request on GitHub, and the app is brought onto the merged code by a merge, never a reset, so nothing Agent committed in the meantime is lost.