AI Toolbox：32美元成本两周上线，两年达成五位数MRR

Scratching his own itch

I'm Adi Leviim, a full-stack developer with 7+ years of building products. I cofounded Infi Developments with Mohammad El-Esawi. Together, we have about 12 years of shipping SaaS, and we split the work cleanly: I lead product development, frontend architecture, and AI integrations, while Mohammad handles backend architecture, algorithms, and LLM systems. Two people, no committee. This allows us to ship a fix the same day a user reports it. In mid-2024, we both used ChatGPT daily for coding, research, writing, and brainstorming. After a few months, we'd each had hundreds of conversations, with no way to find anything within them. No folders. No search within your own messages. No way to export or bulk-manage. We both scrolled sidebars, hunting for a chat we knew was there. So, we built what we 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 as ChatGPT-only, but now it works with Gemini, Claude, and Grok. I'm proudest of the feature nobody asked for: search that spans all four platforms at once. It runs on an index that lives on your own device and never touches our servers. No AI company will ever ship this feature, because building it means indexing their competitors. That's our company's entire thesis in one line. We launched on the Chrome Web Store in September 2024. Less than two years later, we have 35,000+ active users across 150+ countries, a 4.5/5 rating, a Featured badge from Google, and a 5-figure MRR.

Validating and quitting his job

Two developers being irritated by something is not a market, so we looked for other people's complaints before building anything. Specifically, we went where people complain unprompted: Reddit and OpenAI's official community forum. That's a completely different signal from asking people what they want. Nobody writes a forum post about a mild inconvenience. We found the same handful of frustrations repeating in other people's words: I can't find my old chats, there are no folders, I can't search what I actually wrote, I can't get my data out. Those complaints became our first feature list almost verbatim. We weren't guessing what to build; we were transcribing. That made the decision easy, and it wasn't a small one. Mohammad and I were full-time colleagues at the same startup. We met there, and that's why we already knew we could build together. Once we had this idea, we didn't try to squeeze it into nights and weekends. We quit our jobs and went all in on AI Toolbox.

Spending $32 and two weeks on v1.0

The initial build cost us $32 and took two weeks. That isn't a joke or a rounded figure: $27 for a virtual machine to run the backend, and $5 for the Chrome Web Store developer fee. That was the entire capital cost of the company. No agency, no contractors, no investment, no ad spend. Just the two of us and our own time. The first version did exactly one thing: it searched your chat history. That's all. No folders, no export, none of what the product is today. We went through those forum complaints, picked the sharpest one, and built only that. Two weeks after launch, we shipped bulk actions, allowing you to select a pile of chats and delete or archive them, as people kept asking for this next. We now have over 35 features, and almost every one after the first arrived similarly: a customer requested it. The two features people lean on most today — bulk export and the "//" prompt shortcut — weren't in that first build at all.

Architecture decisions

The extension uses TypeScript with no UI framework (no React, Vue, or others). This is a deliberate choice, not laziness: a content script is a guest in someone else's DOM. Every kilobyte we ship adds to the user's page load, and pulling a framework's runtime into a page we don't own is a great way to fight with the host app and lose. We use plain TypeScript and direct DOM work, bundled with webpack, esbuild-loader for speed, and Terser to keep it small. UI that must be bulletproof against the host page renders into shadow DOM, ensuring their styles and ours cannot conflict. Chrome's Manifest V3 shaped our architecture more than any framework decision. Two constraints stand out. First, the service worker can be killed at any moment, so nothing can live in memory and expect to persist. Everything is stateless or database-backed, and every operation must be safely repeatable. Second, and this one bites people: an MV3 content script cannot load JS chunks at runtime. Code splitting silently fails. It's not a build error or a catchable exception; it's simply a feature that doesn't work in production. So we ship one eager bundle and treat bundle size as a real budget rather than something the bundler solves for us. The backend is deliberately boring, and I mean that as a compliment. We use Node and Express in TypeScript, MongoDB, and Redis for rate limiting, all running on a single Ubuntu droplet behind Nginx with PM2 keeping the process alive. That's the whole thing. No Kubernetes, no microservices, no managed anything. The database lives on the same box, which sounds quaint until you realize it means the database is not exposed to the internet at all. We encrypt everything we store at rest. What we store is only the organizing layer: folder names, tags, saved prompts. The one piece of real machinery is PDF export, which runs through headless Chrome on the server — the conversation is rendered in the moment, and nothing is retained afterward. Getting typography and full Unicode right client-side is a losing battle when your users write in Hebrew, Arabic, Japanese, and Chinese. The decision I'd most strongly defend from those early weeks was to run search on a local index. Every conversation is cached in IndexedDB in your own browser, and search runs there, not on our servers. We chose it for privacy, as we were uncomfortable holding other people's conversation history, and it proved to be the best product decision. Search is instant, a bulk export of fifty chats costs zero extra API calls, and two years later, it's the only reason we could build cross-platform search at all.

Shifting the business model gradually

For the first three months, the extension was completely free. Not freemium, not a trial. Free. We had one feature and no reputation, and we couldn't get feedback without users. That period gave us the two things we needed: daily users and a community to tell us what to build next. Almost everything we shipped in year one came from that period. Then, we launched freemium: a free tier with limits and Premium as monthly or lifetime plans, running on Lemon Squeezy. The free tier isn't a crippled demo. It's genuinely useful on its own, which matters because most paying users started as free users who hit a wall doing real work, rather than being sold to. When we went cross-platform, two things changed. We moved billing to Polar and added plans that only make sense once you're on four sites: All Access, which covers ChatGPT, Claude, Gemini, and Grok, and a seat-based tier for teams. That's where expansion lives. A user who buys for ChatGPT and then starts using Claude has an obvious next step; it's the same product they already trust. Every new platform we add makes All Access worth more without changing our build costs, since features are shared modules and each platform is a thin adapter. We were also badly underpriced at the start, and I'd defend that. Early on, we had a handful of features, so we priced it like a product with a handful of features. As the product grew, we raised prices because the product you buy today is not what we sold in 2024. People don't believe our expenses. Our infrastructure bill is a $40 VM. It was $27 at launch, and we upgraded it to something stronger; that upgrade represents our entire infrastructure spend history. The only other running costs are Polar's cut of each sale and about $5 a month of OpenAI API usage for the one feature that needs it: we summarize a conversation for context injection at the moment you ask for it, and we don't store it. We deliberately run it on the cheapest nano-tier model. Two people, one droplet, and $45 a month, at 5-figure MRR. We never raised money, so we never had to grow into a burn rate. This means we can make decisions on a ten-year horizon instead of a runway.