Softgen

The output goes to a repository in your GitHub and a deploy on your own Vercel from the first prompt, so the day you outgrow the builder you already have the code and the URL. That is the daily trait that decides it. What you will not love is that there is no terminal, so when the build breaks you describe the error to the agent instead of reading it, and that gets old by the third bug.

Pick it for an MVP you intend to keep and hand to a developer later. Pick Lovable if you want hosting handled, and Bolt if you want a terminal in the browser.

The risk is verification. The listing shows no terminal execution and no browser tool, so the agent cannot run what it wrote or look at it, and the capabilities page marks front-end, persistence and auth as mature while third-party integrations are still in progress. Integrations judged harmful in the sandbox are not permitted, a limit stated without a list, so you learn the boundary by hitting it.

Assume every generated feature is untested until you test it. What it does right: Stripe and Resend integrations are AI-guided rather than left as an exercise, which is where most prompt-to-app builders stop.

1. Output is a Next.js application. 2. Persistence and authentication go through Supabase by default or a manually configured Firebase, so the data layer is a fixed choice rather than an inferred one. 3. The model is chosen by the user from a list of about a dozen, so results vary with the choice and no two users are testing the same system. No verification step is described: nothing runs, nothing is checked, the output is text that happens to be code.

No benchmark is claimed. A stack, documented; an architecture, not. The observation: the most important variable is the one the user picks from a menu.

In August 2025 Softgen dropped monthly token plans for a $33 a year membership and a pay-as-you-go wallet, which says the subscription churned and the wallet is the retention mechanism now: prepaid balances are harder to walk away from than a cancel button. Advertising usage rates up to 60% below other builders is a race downhill against rivals with more capital and the same models.

No funding signal on record, no distribution beyond search. Likely outcome: a pivot to a niche or absorption into a hosting company. Position: pass, and if you use it, keep the wallet small.

The demo is a front-end from a paragraph. Procurement: sixty memberships is $1,980 a year, which is pleasant, and then sixty separate wallets topped up by card with no central meter, which is not. No enterprise tier, no SSO, no SCIM, no audit log, no stated retention window, and the output lands in accounts we do not administer.

There is no CI story because there is no pipeline; the product is the pipeline. Onboarding is nil and so is control. Not yet, and not until there is an invoice with one line on it.

Nothing to read or run at home, no bring-your-own key, no local models, no MCP in either direction. The tool is a closed box that emits code, and the box is not for sale. What I respect: the model picker reaches past the usual three to Kimi and Qwen, which is more range than most closed builders offer, and the artifact it emits is plain Next.js I can open anywhere.

A fork of the tool is impossible; a fork of its output is a git clone. Rented workshop, owned furniture, which is the only deal a closed builder can offer, and this one honors it.