Two people, one server, $45 a month in costs, and a five-figure MRR.

The companies winning with AI aren't just adopting new tools—they're measuring exactly how those tools are being used across every team.

Discover where AI delivers value, identify potential risks, and gain organization-wide visibility with detailed AI usage insights from Harmonic Security.

The best products often come from a problem the builders felt every single day before they realized it was a market.

Adi Leviim is a full-stack developer with more than seven years of shipping products. In mid-2024, he and his cofounder Mohammad El-Esawi were both using ChatGPT daily for coding, research, and writing — and both drowning in their own history. Hundreds of conversations each, no folders, no way to search their own messages, no way to export anything. They spent their days scrolling sidebars hunting for a chat they knew was in there somewhere.

So they built the fix they wanted: AI Toolbox, a browser extension that adds nested folders, full-text search across every message, bookmarks, a prompt library, and bulk export. It started ChatGPT-only and now works across Gemini, Claude, and Grok as well.

They launched on the Chrome Web Store in September 2024. Less than two years later, AI Toolbox has more than 35,000 active users across 150+ countries, a 4.5-star rating, a Featured badge from Google, and a five-figure MRR — all run by two people on infrastructure that costs $45 a month.

Here is how they did it, including some of the sharpest thinking on platform risk and validation this newsletter has run.

Adi is clear about the trap they avoided: two developers being irritated by something is not a market. So before building anything, they went looking for other people's complaints — and specifically, they went where people complain unprompted.

That meant Reddit and OpenAI's official community forum, which is a completely different signal from asking people what they want. Nobody writes a forum post about a mild inconvenience. The complaints they found repeated the same handful of frustrations in stranger after stranger's words: I cannot find my old chats, there are no folders, I cannot search what I actually wrote, I cannot get my data out.

Those complaints became the first feature list almost verbatim. As Adi puts it, they were not guessing what to build — they were transcribing. That clarity is what made the next decision easy. He and Mohammad were full-time colleagues at the same startup, which is how they already knew they could build together. Rather than squeeze the idea into nights and weekends, they both quit and went all in.

The entire capital cost of the company was $32: $27 for a virtual machine to run the backend, and $5 for the Chrome Web Store developer fee. No agency, no contractors, no investment, no ad spend. Two people and their own time.

Version one did exactly one thing: it searched your chat history. That was all. They went through the forum complaints, picked the sharpest one, and built only that. Two weeks after launch, they shipped bulk actions — select a pile of chats and delete or archive them — because people kept asking for it next.

The product now has more than 35 features, and almost every one after the first arrived the same way: a customer requested it. Tellingly, the two features people lean on most today — bulk export and the "//" prompt shortcut — were not in that first build at all. The roadmap was written by users, one request at a time.

The technical choices are worth understanding because they are unusually disciplined, and Adi defends each one on principle rather than habit.

The extension uses TypeScript with no UI framework — no React, no Vue. This is deliberate. A content script is a guest in someone else's page, and every kilobyte adds to the user's page load. Pulling a framework's runtime into a page you do not own is, in his words, a great way to fight with the host app and lose. They use plain TypeScript and direct DOM work, with UI that must be bulletproof rendered into shadow DOM so their styles and the host page's cannot collide.

Chrome's Manifest V3 shaped the architecture more than any framework decision. The service worker can be killed at any moment, so nothing lives in memory expecting to persist — everything is stateless or database-backed, and every operation is safely repeatable. And because an MV3 content script cannot load JavaScript chunks at runtime, code splitting silently fails in production. So they ship one eager bundle and treat bundle size as a real budget rather than something the tooling solves for them.

The backend is deliberately boring, which he means as a compliment:

The decision Adi defends most strongly from those early weeks was running search on a local index. Every conversation is cached in the browser's own storage, and search runs there, not on their servers. They chose it because they were uncomfortable holding other people's conversation history — and it turned out to be the best product decision they made. Search is instant, a bulk export of fifty chats costs zero extra API calls, and two years later it is the only reason they could build cross-platform search at all.

The real challenge was never the features. It was the platform underneath them. ChatGPT has no public API for your own conversations — no documented endpoint to list your chats, no sanctioned way to read your own history. So the entire product runs on the same undocumented endpoints the web app uses, inside an extension injected into a page the team does not own and cannot predict.

Adi tells one story that captures the whole risk. Gemini redesigned their sidebar. AI Toolbox had a selector that hid Gemini's own search results while showing its own. After the redesign, that selector matched a different element — the sidebar's recent-chats list — and so their CSS hid the user's entire chat history. The data was completely safe. It was just invisible, which is a distinction nobody cares about when they open the app and their chats are gone.

He is candid that this was their fault, not Google's. They had reached for a global selector on a page they do not own. The redesign did not break them so much as find the fragility they had already shipped. That lesson now shapes every decision: assume the ground moves, build multiple fallbacks for every anchor into the host UI, and never let one selector change take the product down.

The feature Adi is proudest of is the one nobody asked for: search that spans ChatGPT, Claude, Gemini, and Grok at once. It runs on an index that lives on the user's own device and never touches their servers.

The strategic logic behind it is the entire company in one line. No AI company will ever ship this feature, because building it would mean indexing their competitors. Each of those companies races on model quality, and none has any reason to build a layer that spans all four. That layer is precisely where AI Toolbox lives.

There is a quiet business advantage buried in this too. Every new platform they add makes their All Access plan worth more without meaningfully changing build costs, because features are shared modules and each platform is a thin adapter. The product compounds in value as the AI landscape fragments further.

For the first three months, the extension was completely free — not freemium, not a trial. They had one feature and no reputation, and they could not get feedback without users. That period gave them the two things they needed: daily users and a community to tell them what to build next. Almost everything shipped in year one came out of it.

Then they launched freemium: a genuinely useful free tier, with Premium as monthly or lifetime plans. The free tier is not a crippled demo, which matters, because most paying users started free, hit a wall doing real work, and upgraded themselves rather than being sold to. When they went cross-platform, they added plans that only make sense once you use several sites — All Access covering all four, plus a seat-based tier for teams. That is where revenue expansion lives: a user who buys for ChatGPT and starts using Claude has an obvious next step in the same product they already trust.

Adi is direct that they were badly underpriced at the start, and he defends it. Early on they had a handful of features, so they priced like a product with a handful of features. As the product grew, they raised prices — because what you buy today is not what they sold in 2024, and they are still raising them.

The expense figures are the part people do not believe. The infrastructure bill is a $40 VM — up from $27 at launch, an upgrade that represents their entire infrastructure spend history. The only other running costs are the payment processor's cut and about $5 a month of API usage for the one feature that needs it. Two people, one droplet, $45 a month, at five-figure MRR. Because they never raised money, they never had to grow into a burn rate — which means they can make decisions on a ten-year horizon instead of a runway.

Adi is refreshingly honest here: they have never had a growth strategy as such. No ads, no growth hacking, no funnel. Their closest thing to a launch tactic was giving the product away for three months.

After that, the Chrome Web Store did most of the work, and it took them a while to understand why. The store is a search engine that ranks on ratings and installs — which means what grows them is not marketing at all. It is shipping what people ask for, which makes them stay, rate it well, and helps the product rank so more people find it. Google's Featured badge came out of the same loop and now drives a significant share of traffic. When someone emails a complaint and they fix it that week, Adi considers that their growth channel, not just support.

Around that core loop, a few deliberate choices:

Adi's obvious advice is to ship something small and read your support inbox. The less obvious pieces are the valuable ones:

The direction is more platforms, more features, more people helped — with one idea underneath all three. Your AI work is scattered across tools that will never talk to each other, and no model company has a reason to build the layer that spans them. AI Toolbox's goal is to be where your AI work lives, whichever model you use that day.

Concretely, that means more LLMs — Copilot and Perplexity are being evaluated next — and more features from their usual source. Adi is refreshingly suspicious of founders who claim a five-year feature vision; he does not have one, the roadmap is public, and most of its content comes from other people's ideas. And he wants more users, because in his framing 35,000 is both a lot and nothing: most people using ChatGPT daily still have hundreds of conversations they will never find again, and no idea a fix exists.

You can look at the extension, visit the site, browse the public roadmap, watch a YouTube playlist of how it works, or join the subreddit.

You can follow Adi on LinkedIn and X, or his cofounder Mohammad El-Esawi on LinkedIn and Instagram.

Leadership is asking: where is AI delivering value for us and where is it creating risk? Right now, most teams have no idea.

With Harmonic Security’s Usage Explorer, you get a complete picture of how your organization actually uses AI, automatically categorized into custom use cases with complete tool-level granularity.

See it in action