Move a Replit App Onto Your Cloud and Your Repo
Replit Agent builds and hosts the app. Production is a repo you control, a database you can restore, and a cutover that does not leave two writers.
By Muhammad Saif and Aashan Khan · Techtribe co-founders
Replit Agent will build the app and host it before you have decided you have a product. That speed is the point of the tool, and it is also the trap. The URL works. A friend can sign up. The database is somewhere you did not choose. The agent can still change production because production, in this setup, is the same place the agent works. Moving is not a vote against Replit. Moving is the moment the app needs a repo you control, a cloud account in your name, and a way to make the agent stop being an author of the live system.
The guide is take your Replit app to production. This article is the cutover: what "already hosted" does not mean, how to leave without two copies taking writes, and which package matches the size of the leave. The incident version, once users are already stuck, is when a Replit or v0 app breaks in production. The Bolt and v0 moves are different traps: Bolt generates the stack in the browser, and v0 delivers the screen. Replit delivers a running host. Plan for the host.
A public URL is not a production system
Replit makes the app reachable, which is more than a ZIP file, and less than an operating practice. Ask a plain question. Who can open the database if the founder is offline. Who rotates a key. Where a backup lives, and when it was last restored onto a second database rather than admired in a settings panel. What happens to in-flight requests if you deploy, and whether "deploy" means an agent run you cannot diff.
If the answers are "the Replit account," you have a single login as a bus factor. That can be acceptable for a prototype you are still ashamed of. It is not acceptable for a customer who has paid, or for data you would have to apologize for. The apology will mention a vendor even when the mistake was a query. You still need the query fixed on a system you can read without waiting for a notebook to boot.
Always-on is part of the confusion. A prototype that sleeps is obviously a prototype. A prototype that stays up looks like a launch. Uptime is not the same as a second environment. If you cannot break a staging copy without breaking customers, you do not have staging. You have production and a sense of caution. Caution does not fork the database.
What you are taking with you
The flow, filmed on the live Replit URL, is the spec. Do not rewrite the product in a slide while you migrate the product. The users who exist, if they are real, are the migration. Count them. Export the rows. Decide which tables are the product and which tables are the agent's experiments. Experimental tables should not get a ride to the new database out of completeness.
The code comes out into a repository you can clone, branch, and revert. A Replit project that only runs in Replit is a dependency, not a codebase. Until a second machine can run the core path against a copy of the data, you have not moved. You have made a backup of a feeling.
Secrets come out last and are rotated, not copied. Anything that has been in a Replit secret, a committed file, or a client bundle is treated as exposed. New keys, in your account, with a spend alert if a model is involved. The old keys revoked the same day, not after you have "confirmed the new app works" for a week of double billing and double access.
What you do not take is the agent's permission to keep shipping. The live system and the editing system have to become different places. That split is the whole reason to leave. Recreating the split poorly, by giving the agent your production cloud credentials so it "feels the same," recreates Replit in a more expensive room.
One writer, then the domain
Freeze writes on a date you tell anyone who has data in the app. Deploy the new host. Prove the core path against the imported copy: a second user, a row they may see, a row they may not, a payment if payments exist. Then switch the domain. Then make the Replit app read-only or turn it off.
Any row written to Replit after the freeze is a row you must move by hand or abandon in writing. Two live apps will both accept signups if you leave both open, and you will not merge them with a script you invent on Sunday. The Bolt guide says the same thing about a fallback window. It is stricter here because Replit was already the real URL. People have bookmarks. Bookmarks are why the domain cutover is a moment, not a vibe.
If you have no real users, you can skip the tenderness. Export nothing you do not believe. Turn the old app off when the new path works for you. Do not import sample rows "just in case" and then build permissions around garbage.
A rollback is a deploy of the previous release on your host, practiced once. It is not "turn Replit back on and hope the data is close." If you might need Replit as a fallback, the fallback has to be read-only, or you have two writers again. Pick. The recovery order when a deploy has already gone wrong is in how to recover an AI-built app. Practice it on the new host before the domain moves, not during the apology.
Auth and data, again, because the host change will not do them
Moving servers does not create a tenant boundary. If the Replit app returned every row to the browser, the new host will too, until someone changes the query. The move is the right time to change it, because you are already touching the path. Shipping the same leak on a more respectable domain is a common outcome and a bad one. Two accounts. The test is the launch gate. Repair or rewrite is how you decide whether that gate is a patch on the exported code or a new core.
Money moves with the same skepticism. A payment button that worked because Replit's environment had a webhook pointed at a dev URL will fail, or double, on the new host. Point the webhook at the new URL, verify the signature, and make the handler safe to retry. If you do not yet have a ledger, do not migrate a button. Scope the subscription as its own decision, in add subscriptions to an AI prototype.
Background work is the piece teams forget because the Agent hid it. Scheduled jobs, replays, a bot that posts to Slack from the Replit box. Inventory them before you turn the old box off. A job that only exists in a Replit cron is a feature. It needs a home on the new host or an explicit death. Discovering it from a customer on Monday is the expensive way to inventory.
What the agent is allowed to do later
After the cutover, Replit can still draft a change. The draft arrives as a diff in your repo. A person merges it. The agent does not hold the production database URL. It does not hold the payment key. It does not run against production "just to see." The failure mode you are avoiding is an agent that deletes or rewrites live data because the environment it could see was the environment customers use. That failure is not theoretical in the industry. The permissions article, Claude Code and Cursor can reach your files, is the general version. Your version is shorter: the agent is not an operator.
If nobody on the team can read a diff, do not keep an agent in the loop as a substitute for that person. An AI product your team can maintain is the standard. One safe change a non-engineer can make. Auth and payments stay with an engineer. A Replit login shared in a password manager is not that standard.
Which package is the move
A single path, extracted and stood up for one user, with no second tenant and no billing, can be a short build, or smaller if the extract is truly a repair. Do not book a short engagement and expect a faithful copy of every Replit sidebar. The sidebar was free to generate and expensive to mean. Say that on the call if the sidebar is what you are attached to. We will tell you to leave it behind.
The usual shape of "we have users, or we are about to" is a longer fixed build. Your cloud, your repo, auth that holds for a second person, billing hooks if money is real, a few integrations, a rollback you have used, and monitoring if a model is on the path. Replit will have wired whatever the prompt mentioned. Mentioned is not the same as in-scope. Name the ones the core path needs. The rest wait. We count them with you in the meeting. We will not quietly expand the list afterward.
A team that stays is for the product that keeps going after the domain moves. The move itself should stay a fixed scope so it ends. An open-ended migration is how a team gets paid to reenact the agent. Hire a dedicated AI team describes the pod. Book that when the roadmap is real. Book the fixed scope when the job is the leaving. The call is where those two get separated.
After you leave Replit, the host, the database, and the model are yours to operate. Replit's convenience was hiding some of that. The move makes it visible. Visible is what you wanted. Mention it on the call so the cutover is not a surprise about who holds the accounts. The build and the cloud account are different conversations, and we will keep them separate.
The first week after the domain moves
Expect a short list of things the agent used to do that nobody wrote down. A header on an outbound email. A file path that only existed on the Replit disk. A redirect someone bookmarked with the old host in it. Budget a few days of those, inside the support window, rather than pretending a move is invisible. Invisible moves are the ones that forgot a cron.
Do not invite the agent back to "clean up" that list against production. The cleanup is a commit a person can revert. If a job was Replit-only and you killed it on purpose, write the death in the handover so a helpful teammate does not recreate it. Resurrection of a forgotten bot is how a second writer returns.
Watch the processor and the database side by side for the first renewals or the first real signups. One row each. If they diverge, stop inviting people until the handler is unique again. Growth on a double-writing webhook is not traction. It is a queue of apologies. The same watch applies if you had no payments: one signup, one account, one set of rows that account may see, and a second account that cannot.
Tell anyone who had the old URL what changed, in one sentence. Include the date, the new place to sign in, and whether they need to do anything. Most of them need to do nothing, and saying nothing is the courtesy. The ones who do need to act, because a bookmark or a webhook they own still points at Replit, get that sentence separately so it is not buried under reassurance. The product is the same job, the host is yours, and the day of the freeze is when data stopped being accepted on the old copy. People forgive a move they were told about. They do not forgive a form that saved into a database you already retired.
What to bring to the call
The Replit URL. A note on whether strangers, paying users, or only you have accounts. An export of the schema, or access for us to read it. The list of crons, webhooks, and third-party tools. The one flow that has to work on the day of the cutover, and the pages that can stay dark.
If the app is already broken, bring the break, not only the dream of a clean cloud. The debug path is how to debug a vibe-coded app: one path, the real error, then the fix. A migration that begins during an outage, with no freeze, is how data splits. Stabilize the writer you have, or freeze it, then move.
If this is your app, send the repo.
A 30-minute scope call. We read what is live, name what is dangerous, and quote a fixed scope.
How Techtribe moves a Replit app
We export the code into your repo and the rows you intend to keep. We stand the core path up on your cloud, with staging that is not the production database. Two accounts prove the access rule. Webhooks and jobs are named, moved, or killed. Secrets are rotated. The domain moves when there is one writer. Replit Agent does not receive production credentials afterward.
A small extract can stay a repair. A cutover people will rely on is a fixed build with an end. The month after a real roadmap is a team that stays. We say which on the call. The guide is Replit to production. If the app is already failing users, read when a Replit app breaks in production and bring the failure to the call. The service page is AI MVP development.
The scope call is 30 minutes with a founder. Bring the live URL and say whether the data is real. That is enough for us to tell you if you should move, and which parts of the app come with you.
Questions
If the Replit app is already online, why move it?
Online on Replit's host is a preview that strangers can reach. It is not a repository you control, a database you have restored, or a deploy you can roll back without the agent. Those are the production step.
Can Replit Agent keep editing the app after we move?
Only inside a boundary you write down. Low-risk changes, on a branch, reviewed by a person. The agent does not get production credentials, and it does not edit auth, payments, or the data rules.
What has to be true on the day the domain moves?
One writer. The new host serves the core path. The Replit copy is read-only or off. Any row created after the freeze has a plan, because two databases will not merge themselves.
How do we know the move is the right size?
Book the scope call and bring the live URL. A small extract of one working path stays small. A second user, your cloud, and a rollback you have actually used is the move people rely on. We say which after we know whether the data is real.