Bolt AI Review

The setup

I signed up for a fresh free account with no card attached and gave myself one rule for the test. Every screenshot would include the token counter, because the counter is the part of Bolt that nobody photographs and everybody argues about.

The free plan allows 300,000 tokens a day against a monthly ceiling of one million. My question was simple. How far does a day get you on one app?

The test project was ShelfTrack, a home inventory tracker. Small enough to finish in an afternoon, complicated enough to need a login screen and a real database behind it.

The first build

I opened Bolt with a full balance and pasted a single prompt describing the whole app.

Generation was fast. The file tree filled in, npm pulled packages down in the terminal, and a live preview rendered before I had finished reading Bolt's own description of what it was doing. This is the WebContainer running Node inside the browser tab, and the speed advantage is not marketing.

The login screen came out clean, with a note that demo mode would accept any credentials. That was expected, since I had asked for mock data.

The dashboard arrived populated. Twelve item types across ten locations, forty-six units in total. Search filtered the table. The Add Item modal opened and closed correctly. Version 1 was stamped 1:28 PM.

The counter read 200,000. One prompt had taken a third of the day.

The database was not included

The app looked finished. It was not. I opened the database panel and found an empty pitch for a backend I had assumed came with the build.

Nothing here is automatic. You either ask Bolt to create a database or you bring Supabase yourself. I asked Bolt.

Setup finished at 1:50 PM. Type checking and the production build both passed, which I verified rather than took on trust. The balance had fallen to 105,000.

Ninety-five thousand tokens for one table.

A padding tweak cost 65,000 tokens

For the next prompt I asked for the smallest change I could think of.

Look at what Bolt logged. Three actions, all against Dashboard.tsx. It read the file, edited it, then compiled to confirm nothing had broken. No new components. No dependency changes.

That cost 65,000 tokens, which is twenty-two percent of a free day spent on row spacing.

The explanation sits in the architecture rather than in the size of the edit. Bolt synchronises the full project tree to the model on every prompt, so the bill tracks how much code exists rather than how much of it you asked to change. A cosmetic tweak on a twelve-file project already costs this much. The same tweak on a fifty-file project costs more, and there is no setting that opts you out of it.

The inventory that disappeared

With the database live, I registered a real account and signed in.

My first assumption was that Bolt had wiped my sample data. It had not, and the distinction matters. The demo inventory had been mock data living in the frontend. Once a real backend took over, my freshly registered account was legitimately empty. The database panel confirmed which of the two had happened.

Bolt moved the app onto a real backend without seeding it and without warning that the inventory on screen was about to vanish. Its summary message told me data now persists across refreshes and stays separated between user accounts. Both claims are accurate. Neither mentions that the dashboard I had been looking at for twenty minutes was about to empty itself.

One more thing sits in that panel. It says to ask Bolt to create or modify tables. There is no manual schema editor. Every column you add later goes through the AI and costs tokens, where a normal Postgres dashboard would let you add one for free in four seconds.

The code underneath is fine

This is the part of the review I expected to be harsher about. The generated code holds up. TypeScript throughout, proper hooks, loading and error states, a signin and signup mode toggle, validation before submit. A React developer could open this and keep building without wanting to start over.

Two details in the file tree still bother me. The file mockData.ts is still sitting there after the migration, unused and unreferenced, left behind as dead weight. And src/ is flat, seven loose files with no components directory.

At this size that structure is harmless. It is also the shape that gets expensive later, because every one of those files gets re-read on every prompt.

Where the day ended

The full accounting from a single afternoon on the free plan:

Two hundred and sixty thousand tokens. What I had to show for them was a login screen, a dashboard, an empty table and slightly tighter padding.

The daily cap arrived before the app had a single row of real data in it.