Money-Making in Action · Episode 21: Five AI Toolchain Paths to Profit, from Zero Barrier to Advanced

Li Bo, this week I was browsing the GitHub trending page and found something weird. A ton of projects with hundreds of thousands of stars, all built to serve agents — but where's the money? You're not wrong, Xiaoyu. But that's not the point. The point is, the more competitive the tools get, the more the shovel sellers make. I pulled three all-nighters and put together five paths that actually work. Five? Well, so much for leaving work early today. Start with the ones a beginner can pick up — the kind that even I'd be tempted by. Number one, the easiest. You remember that video? The guy who can't code making thirty grand a month — he was using tools like Trae, from chatting to shipping a whole system. I remember — the nail salon booking thing. So what's the core of this path? Is it really zero barrier? The target audience is people who can't code but can land clients. The toolchain is Trae plus Claude — use Trae for requirements and solutions, use Claude to finalize the code. The steps are clear. First ask about budget and occupation to judge whether you can take the gig. Then throw the client's one-sentence request at the AI and have it produce a flowchart and technical plan. Wait — the key here is the AI helps you question the client, right? It even generates the talking points? Exactly. It turns open questions into a questionnaire, and the scripts are ready to copy-paste. After the conversation, feed the feedback back in, it breaks down the task list, and in seven to ten minutes you've got a working system. Low difficulty, almost zero cost. ROI depends on order volume — list a small tool online and orders keep coming. The key is you can pause and adjust at every step. I buy that. But here's the question — why would a client trust someone who doesn't know tech? Isn't that a delivery-trust problem? Good question. The answer is path number two: turn the agent into a senior engineer. Addy Osmani's Agent Skills — a hundred thousand stars, twenty-five skills plus nine slash commands. The core is verification gates. The agent can't just say it's done — it has to hand over test runs and diffs. That locks in your delivery quality. So this one is insurance for the first? So even an outsider can deliver to an engineer's standard? Exactly. It's compatible with Claude Code, Cursor, Codex, Gemini CLI. Install it and your order pipeline gets a QA step. I recommend it to anyone trying to make money taking orders. Okay, now let's get into something more technical. I keep seeing headroom lately — the token compression thing. Can that make money? Path number three: saving tokens is saving real cash. headroom compresses tool output and logs before they hit the LLM — twenty percent savings for coding agents, up to ninety-five percent for JSON. The target is teams running batch agents. Toolchain is headroom plus Claude Code, deployed as a proxy or MCP server — logs pass through before hitting the model. I'm sold on this one. It's like putting a water-saving valve on the pipe — heavy-usage teams can cut a big chunk off their monthly bill. Medium difficulty, you need some engineering sense. ROI shows up right on the bill — especially in JSON-heavy scenarios, what you save is profit. This is selling a cost-saving solution. Then there's one I really want to hear about — running several models at once. That OpenCognit thing. Path number four: multi-agent parallelism. One central controller runs Claude Code, Codex, and Gemini CLI at the same time, dividing work and cross-checking. The scenario is very real. One model generates, another reviews, a third checks docs. Split a big task and run it in parallel, and your delivery cycle shrinks directly. So it's basically using several models as a small team? Then I can take big orders solo and still have team-level output. That's the idea. There are pitfalls with resource consumption and result integration, but difficulty is medium-high. ROI is doubled capacity — good for people taking on complex projects. But with so many models, how do you manage them? Each locks its own config — wouldn't switching back and forth be exhausting? Path number five is called Magpie. A menu bar app — one list lets you assign a model to each agent. Claude Code runs Kimi, Codex runs DeepSeek, switch anytime. It also auto-switches on quota. When one account runs out, it quietly swaps to a backup — the agent never notices. Cross-platform across three, MIT open source, sixteen thousand stars. This is infrastructure. Others sell services on top, and it collects the toll. Right. Low difficulty, but the business opportunity is in hosting and configuration services around it. ROI is long-term compounding — good for people running tool sites. Let me confirm one more — there's also llm-course, the one with eighty thousand stars. Is that a money-making path too? It's the foundation, not the solution. The target is people wanting to transition — a roadmap plus Colab notebooks. Run through it and you've got the base to take on those five paths. So to sum up: from taking orders to QA, to saving tokens, to multi-model collaboration, to model switching — a complete money-making pipeline. That's the conclusion. Tools are no longer competing on model capability — they're competing on workflow and quality control. Whoever gets the process running smoothly gets paid. Alright, I'm going to try that nail salon idea tonight. Listeners, don't just listen — pick one and get started tonight.