What Replit really does
Opens a real browser. Clicks and types. Can log in. Sends a summary. Records video. Then fixes what it found. This is the honest counterexample to "nobody actually runs the app." Replit does.
For Replit Agent
Replit's AI does test your app. It also marks its own exam.
We are not knocking Agent 3. It really does open a real browser, click around, type into forms, log in, and send you a summary and video. But then that same AI decides whether what it found is good enough, and fixes it. That is not an independent decision about whether to launch.
Last verified 18 August 2026. Quotes are Replit's own words. The conclusion about independence is ours, and it is our opinion, not theirs.
Tests the apps it builds, using a real browser. Checks buttons, forms, and data. Replies with a summary of its tests, and fixes any issues that it found.
It will periodically decide to test itself, when it thinks enough has changed to be worth checking. That is Replit's own description of this self-testing feature.
Sticking to what Replit itself says
What Agent 3 actually does
Replit says its Agent 3 tests the apps it builds using a real browser, checking buttons, forms, and data, "ensuring every part of your app works perfectly," then replies with a summary and fixes anything it finds. Replit's engineering team describes it as built to catch fake-looking screens that don't actually work, and markets it as faster and cheaper than having a separate AI drive the browser.
The technical detailReplit's blog describes a REPL-based verification system built to catch "Potemkin interfaces," positioned as faster and cheaper than Computer Use models.
Replit's own documentation adds the limits they publish themselves: the agent only decides to test itself sometimes, when it thinks enough has changed; it only works for certain kinds of apps; and a cheaper "Lite" mode turns testing off entirely. What you get back is a written summary and video. We have not watched those videos ourselves, and we are not claiming they are fake. That would be a claim we cannot back up.
The technical detailApp Testing is limited to Full Stack JavaScript and Streamlit Python web applications; Lite mode keeps it off.
Cited: Agent 3 launch, automated self-testing, App Testing docs. PROMPT_TO_APP_PIPELINES Part 2.
Our conclusion, not something Replit admits
Grading your own homework can't produce an honest "we don't know."
What it can't be
An independent examiner. Whoever wrote the code also chooses when to take the test, writes the questions, marks the paper, and edits its own answers. A marketing claim that everything works perfectly can't coexist with an honest "we don't know yet." That conclusion is our read of their own language, not something Replit said about itself.
A link that keeps moving
Your .replit.dev link is public, but it doesn't stay put.
The working link you get while editing in Replit is public by default, but it only works while you are actively in the workspace. It can change every time you reopen the app, and you can make it private at any time. A published app on .replit.app is the stable, real link.
The technical detailDevelopment URLs look like UUID.servername.replit.dev. The owner can flip a Private development URL setting, which requires authentication. Deployments on .replit.app are the stable real-server path; Autoscale can scale to zero.
So a Replit working link is not something you can rely on staying the same: it may have gone dead, changed, or been made private. Do not treat an error there as proof "the app is broken."
Cited: Replit development URLs. TARGET_ACCESS §3.1, §3.3.
How to bring a Replit app
Use Replit's Git tab. It's a normal code repository underneath.
Replit's built-in Git tool can set up a repository, connect it to GitHub, GitLab, or Bitbucket, and handle the usual save-and-upload steps. What comes out the other end is just a normal Git repository, the same kind developers use everywhere. Sign in to PeerRun and bring that remote, or a copy on your disk.
The technical detailReplit's Git interface can initialize a repo, connect GitHub, GitLab or Bitbucket, stage, commit, push, branch, and resolve conflicts. Shell git commands stay in sync. Self-hosted providers are on Enterprise.
If your app depends on secrets Replit hosts for you, or an outside database you have not exported, PeerRun will check what it can and mark the rest BLOCKED or UNKNOWN, and say why. We do not have a partnership with Replit.
Cited: Git interface. TARGET_ACCESS §5.7.
- Signed-in product
- Unknown until you grant access. Agent 3 can log in because it lives inside your workspace already. We do not inherit that session.
- A bot wall
- Blocked, not broken. We have no way around it.
- A dead working link
- Not evidence the product failed. That link may simply have gone away when you closed the workspace.
- What we do not claim
- That Replit's tests are fake, unused, or always pass. That we have watched their videos and found something wrong. That we are certifying anything based on their recordings.
Questions people actually ask
Replit, specifically
Agent 3 said every part works perfectly. Why run PeerRun too?
Because the builder is still the one grading itself. A second opinion is the whole point. We take your export or repo and reach a verdict the AI that wrote your app cannot hand itself.
I only use Lite mode. Did Agent 3 even test my app?
Replit's own docs say Lite mode turns App Testing off. If you never saw a test summary, there may not have been a browser test at all. That is Replit's published limit, not something we measured ourselves.
Is my .replit.app link enough on its own?
It is the stable, working link. But it still only proves what someone who is not logged in can see, unless you give us more access.
Sources
What this page is allowed to say
Let Replit keep running the browser it already runs. Let someone else mark the paper.
Sign in and bring the GitHub, GitLab, or Bitbucket repo from Replit's Git tab.
We will not invent a pass. Anything behind your sign-in stays Unknown until you grant access. A robot-check is Blocked, not broken. The first look is on us.