Everyone asks the same question: how do you actually make money with AI? Not in theory. Not with a pitch deck. With real code, real users, and real revenue.
Ariel’s answer in this video is one project, one developer, and a lot of patience. Meet LocalGPT: a tiny OpenClaw alternative built in Rust over four nights, now evolved into something far more interesting, and quietly earning by turning the app into a lead factory for services and projects.
The origin: four nights in Rust
The story starts on February 26, 2024, right when OpenClaw had just exploded. OpenClaw showed the world that an AI could be given a goal and left to figure things out. It was also a dependency monster: 460,000 lines of code, hard to install, hard to run, and overkill for many use cases.
Then one developer looked at OpenClaw and stripped it to the bone. His insight: you do not need all seven old orchestration markdown files. You need the agent’s soul.
OpenClaw gives each agent seven markdown files to keep it coherent:
- SOUL.md โ personality and tone.
- USER.md โ who is helping whom.
- PROTOCOLS.md โ rules that always bind.
- AGENTS.md โ agent-specific notes and tools.
- MEMORY.md โ what it must remember long-term.
- IDENTITY.md โ who the agent is.
- HEARTBEAT.md โ what it should keep doing.
LocalGPT trimmed that down to a lighter set. According to the project docs, the main files are soul.md, memory.md, heartbeat.md, policy.md, LocalGPT.md, and AGENTS.md for tool discovery. Less ceremony, less orchestration, same soul.
Four nights of Rust later, LocalGPT compiled into a single executable. Fifteen thousand lines versus OpenClaw’s 460,000. Plug in any model, local or remote. Run it in a terminal. Done.
Ariel even found his own tiny commit in the project history. He had assumed LocalGPT was dead. It was not.
From terminal toy to 3D agent
The project kept growing. The developer added a world agent skill: the agent can connect to a game engine and let you sketch shapes, terrain, and areas in a browser-based 3D editor. Pyramids, spaceships, whatever you describe.
Then came quark, a desktop window with the familiar controls: model selector, assistant threads, the usual AI helper layout. LocalGPT went from a single-file CLI to a small ecosystem without ever becoming bloated.
That is the trick. It stayed small enough to understand, but capable enough to keep users coming back.
How the money actually works
Open source does not pay by itself. LocalGPT’s monetization is built so the app itself becomes a lead factory for services and projects:
- “Get in touch” โ a form that asks about budget, infrastructure, users, model needs, compliance, and timeline. In other words, a direct pipeline for paid consulting and custom deployments.
- Projects and services โ once people use the tool and see value, they come back asking for custom installation, integration, or hand-holding inside their company.
Ariel compares this to Zig, which funds a small team from open donations, and Ollama, which is open source but sells hosted Ollama instances and support. The pattern is old: give away the core, charge for convenience, expertise, and scale. For LocalGPT, the real income comes from customers asking for professional help โ not from a sponsorship button.
The real lesson
If you build something genuinely useful, do not obsess over the first dollar. Help people first. Make the tool good. Then add a paid tier, a “contact us” button, or a support contract. Ariel says the same thing to a community member building an education product: take your time, build a good SaaS, offer a free part and a paid addition.
LocalGPT has been in development since February 2024. That is nearly two years of steady investment before the money becomes real. There are no overnight successes in serious open-source tooling.
๐ฅ Roast Corner
Let us start with the obvious: this entire success story depends on one person who can write Rust in four nights. If that is not you, LocalGPT is not a business plan, it is a hobby with a contact form.
And the “Get in touch” form that interrogates you about budget, compliance, and timeline before saying hello? That is not a feature, that is a sales funnel wearing a lab coat. It works because the project is good, not because the funnel is revolutionary. Do not copy the wording. Copy the two years of grinding.
Finally, the pigtails. We are not going to discuss the pigtails.
๐ค AI for Humans
Imagine you have a small team that wants local AI agents running on their own machines, not in someone else’s cloud. You could hand them OpenClaw and watch them drown in dependencies and seven markdown files. Or you could hand them LocalGPT: one file, one config, one model endpoint. They get 80% of the power with 10% of the headache.
That is the business model hiding inside the code. Not replacing OpenClaw. Serving the user who wants power without the forklift.
If you are sitting on a side project right now, the lesson is simple: ship something small and useful, keep showing up, and let the money arrive after the users do.
๐ฌ Comments