How we review and compare AI tools

Every review and comparison on this site follows the same method, so you can judge our conclusions instead of trusting them. This page is the method. Who we are and how the site is paid for is on the about page; who writes it is on the author page.

What we do and don't do

• We compare published facts. Prices, plan limits and features come from each vendor's own pricing page, documentation or help centre, opened on the date shown next to them, with a link.
• We say when we've used something. Mahmoud has built with Lovable and uses Claude Code regularly, and those reviews say so in a box at the top. Everywhere else, the review is from published specs and docs, and says that.
• We don't run synthetic benchmarks, publish other people's test results as ours, quote anonymous users, or invent ratings.
• We don't take money from the products we review. No affiliate links, sponsorships or paid placement.

The rubric

Each product gets a score from 1 to 5 on six criteria, with a one-sentence reason shown next to every score. The overall score is the plain average, rounded to one decimal. We use the same six criteria for app builders and coding agents on purpose: they're alternatives for many of the same jobs, and the rubric shows where each kind is strong.

The criteria are equally weighted because the right weights depend on you. A founder testing an idea should care most about speed and backend; a team with a codebase about depth, exit and team. That's why every review leads with "pick it if" and "skip it if", not the number.

Scores are judgements, made by one person against published facts, and two people could reasonably differ by a point. They're most useful for comparing products on the same criterion, and least useful as a single ranking.

Where the data comes from

All product facts live in one dataset that every review, head-to-head, table and the home page read from, so a price is the same wherever it appears. Each price carries its source link and the date we checked it ("checked Nov 1, 2026"). Where a vendor's own pages disagree, we say so on the page (Replit's annual Core price is one example). Where a vendor doesn't publish something, such as usage quotas, we say it isn't published rather than estimating it.

For agent frameworks, we use each project's documentation and GitHub repository; star counts are a popularity signal, dated, and never part of a score. For security guidance, we cite the MCP specification, OWASP and the original incident write-ups.

How often we check

Prices in this market change monthly. We recheck every price and plan on the site at least once a month and after any announced pricing change. When a fact changes we update the page, change its "updated" date and note it on the updates page. We don't change dates when nothing changed.

Our own products

We run Skillspace (skillspace.sh), a directory that grades public agent skills. It isn't an app builder or coding agent, so it isn't scored in the reviews table. Where it does appear, on the registry comparison, the MCP and skill vetting guide and the pattern library, it's labelled as ours, its weaknesses are listed alongside everyone else's, and it isn't ranked first by default. We never use Review or rating markup for a product we own.

Mahmoud's studio, Space & Story, builds software for clients. It's mentioned once, on the app builders comparison, as an option for when a project outgrows the builders, with that interest stated.

Corrections

If something is wrong or out of date, tell us through mahmoudhalat.com. We fix factual errors as soon as we can confirm them, update the page date, and list the correction on the updates page. Vendors can send corrections the same way; we'll change a fact when a public source supports it, and we don't change scores or verdicts in exchange for anything.