Starts.live

Inspiration

I've spent four years building Toolblox and, before that, over a decade shipping products for clients. After building hundreds of MVPs, one pattern was impossible to ignore: the code was almost never the problem. The market was. CB Insights puts "no market need" at the top of the startup killer list, and I'd watched it happen up close — months of clean engineering for products nobody had asked for.

AI made this worse, not better. Tools like Lovable, Bolt and Cursor will happily build whatever you describe, fast. You get exactly what you asked for, in a complete echo chamber. Venture studios solved this decades ago — pressure-test first, build second, but they take 30-50% equity, accept 1-2% of applicants, and mostly want you in San Francisco. I wanted that playbook available to anyone with an idea and a laptop.

What it does

You describe your idea in one sentence. Within an hour, Starts.live has a real branded landing page live on its own link not a mockup, with multiple headline variants A/B testing against each other from day one. Within a day, it's running real ad campaigns on Meta and Reddit, and sending personalized cold outreach to verified contacts. Within three days, you have a verdict backed by real numbers: signups, replies, CTR, scroll depth, cost per click.

Then comes the part most validation tools skip: it builds. The AI team reads the data, doubles down on what pulled, drops what didn't, and ships actual product features — login, payments, dashboards — inside a private per-project sandbox with its own database, GitHub integration and version history. You direct everything from chat. You keep 100% of your company.

How I built it

The platform runs each project in an isolated sandbox with its own React + Vite workspace, database, and controlled internet access, so the agent can install packages and ship real features without projects seeing each other. The agents run on the Claude model family, with heavier models reserved for hard reasoning and faster ones for routine loops. Ads go out through the Meta and Reddit ad platforms; outreach runs on Hunter for contact discovery and verification plus Amazon SES for delivery. The marketing brain isn't improvised — the copy, positioning and test sequencing encode Lean Startup, April Dunford's positioning work, Cialdini and Jobs-to-be-Done as explicit method, applied automatically.

Challenges I ran into

The hardest part wasn't generating pages or copy, AI does that easily now. It was closing the loop safely with the real world. Real ad accounts, real spend, real emails to real people means the agent's mistakes cost money and reputation, so every external action needed guardrails, budget limits and review gates. Ad platform policies and cold-outreach deliverability are moving targets that punish naive automation. And sandboxing was non-negotiable: giving an agent real internet access and package installation while keeping every project sealed from every other one took real architectural work.

Accomplishments that I'm proud of

Real projects have gone through the full loop — live pages, live ads, real strangers signing up — and the platform currently shows 147 projects live. The proof section on the site links to actual ads and landing pages the system built, because "proof, not promises" only works if you can click on it. Mostly I'm proud that the thing does what venture studios charge six figures and a third of your company for, and it does it for a monthly subscription, solo-bootstrapped from Tallinn.

What I learned

That the venture studio insight was right all along — the market has to be in the loop from hour one — and that the work studios do is mostly repeatable method, which is exactly what agents are good at. The founder shouldn't be automated; everything between the founder's idea and the market's answer should be. I also learned that agent autonomy is an economics problem as much as a technical one: the moment an agent spends real money, the interesting engineering is in the constraints, not the capabilities.

What's next for Starts.live

Deeper channel coverage beyond Meta and Reddit, richer signal from the outreach side, and pushing the build loop further so the distance from "the data says customers want X" to "X is live and being retested" keeps shrinking. The goal hasn't changed: anyone with an idea gets a real market verdict in three days and a product built with the customers it found — while keeping all of their company.

Built With

• claude
• next.js