Emergent AI app builder Review

I ran Emergent on its free tier, which starts you with 10 credits and asks for no card. My goal was a single build session that would surface both the good and the bad, so I asked it for one small, familiar app: a restaurant bill splitter aimed at US diners. One prompt, a few follow-ups, and a close look at what actually shipped.

The starting prompt was plain English: build a web app called SplitEasy for splitting restaurant bills, with a USD total, tip presets at 15%, 18%, and 20%, a diner count from 2 to 10, and a per-person figure. The credit meter read 10.00 before I sent it.

The build: agents that narrate their own work

Emergent did not just start typing code. Within seconds it split the job across named agents and showed each one working. Before committing to a look, it stopped and asked me a design question, which is closer to how a real teammate behaves than the fire-and-forget generation I expected. I picked the "surprise me" option to see what it would choose on its own.

The part that stood out: the agent took its own screenshots of the running app and read them back to verify the build matched the prompt. Watching it capture the UI, inspect it, then move on felt like genuine self-checking rather than a progress bar dressed up as one.

It then handed the work to a testing subagent. The transcript is confident about the result. It states the bill math is correct, works a sample split of $156.80 at 20% across four people, and reports 15 of 16 checks passing, dismissing the one failure as a flaky test script rather than an app bug. Hold onto that claim, because the finished app tells a different story.

Speed and the finished screen

Timing was the easy win. My prompt went in at 12:44 PM and the agent posted "Agent Finished" at 12:52 PM, so a complete full-stack app landed in eight minutes. The right panel showed a working preview and a Manage tab with a Payments section offering a Stripe hookup, useful for any US reader who wants to charge for what they build.

Where it breaks: the one number that matters

SplitEasy exists to answer a single question. How much does each person owe? That is the number it gets wrong.

I opened the app in a new tab and entered a $670 bill for 2 diners at 0% tip. The receipt panel computed the subtotal, the tip, and the total perfectly, printing $670.00. Directly below its own correct total, under "each person pays," it showed $3.54. The right answer is $335.00.

I assumed I had triggered some odd edge case, so I tested a clean one. A $100 bill at 20% tip across 4 diners. The receipt again computed everything correctly, showing a $120.00 total, then printed the same $3.54 per person. The right answer is $30.00.

Two unrelated inputs producing the exact same wrong figure means the app is not miscalculating the split. It is not calculating it at all. The $3.54 is fixed no matter what you type. Every other number on the receipt is right, which makes the failure worse, because the app looks trustworthy until you check the one value you came for. This is also the number the testing subagent swore it had validated.

The free-tier ceiling: credits and code

Cost adds up quietly. One app with a couple of iterations moved the meter from 10.00 to 6.40 to 4.42, roughly 5.6 credits gone in a single sitting. On the free allowance, a second serious attempt at the same app would run me close to empty.

I also went looking for the source code and could not open it. The agent clearly wrote real files, since the transcript references paths like /app/memory/PRD.md, but inspecting or exporting that code sits behind the paid tiers. On free, you get a working preview and no way to read what produced it.

Editing, at least, is open. The preview carries an inline "describe a change" box and an Edit control, so you can reshape the app in place with more prompts without touching a settings menu.

Emergent delivered a designed, deployable app from one sentence in eight minutes, complete with agents that question you, screenshot themselves, and run their own tests. It also shipped a bill splitter that cannot split a bill, then certified the math on the way out the door. The presentation is strong enough that a less suspicious user would have published the broken number without ever noticing.