By Katsushi Sakuma · Published 2026-07-17 · Last updated July 17, 2026 · 3 min read
Every open-source author eventually hits the same wall: your code is free, your time is not. Donations don't scale, dual licensing scares users away, and SaaS-ifying your project means you're no longer really doing open source.
I think x402 finally gives OSS authors a fourth option. I call it OSS body, metered brain, and I've been running it in production with my crypto trading system, Kurage FreqAI Trading System (kfreqai).
The split
kfreqai is an AI crypto auto-trading system built on Freqtrade/FreqAI, licensed GPLv3. Anyone can clone it and run the full trading engine - exchange connectivity, ML pipeline, risk gates, dashboard. That's the body, and it's genuinely free.
What the repo deliberately doesn't ship is the judgment layer: market-regime classification, daily risk directives, per-trade post-mortems, strategy-hypothesis generation. Nine LLM-driven functions in total. That's the brain, and it's exposed as a paid API behind x402, $0.05 per call, settled in USDC on Base.
The key design decision: the brain is a separate program that never touches Freqtrade's code. It talks to the body over REST, holds no exchange credentials, and cannot place orders. This does two things at once. Legally, it keeps the paid layer independent of the GPL body. It's not a derivative work, so I'm free to license it commercially. Operationally, it's a trust boundary: even a malicious brain response can't move funds, because the final risk gates live in the OSS body the user runs themselves.
Why x402 and not API keys
The users this serves aren't humans filling out signup forms. They're agents and scripts. A fresh kfreqai clone runs free on a rule-based fallback brain. Set one environment variable and it starts paying per call for the real one. No account, no subscription, no credit card form: the 402 challenge is the onboarding. For an OSS project whose users are, definitionally, other people's software, that's the whole funnel.
Micropayments also match how trading bots actually consume judgment: a regime call every hour, a directive once a day, a post-mortem per closed trade. Nobody wants a $50/month plan for that; five cents at the moment of use is the honest price.
What I learned running it
Match the paid engine to its terms of service. Locally, my judgment functions can ride a tiered fallback (Claude CLI → Codex → local Gemma). The paid x402 path forces local Gemma 4 12B only. Reselling a subscription CLI through a paid API would violate its terms. If you adopt this model, audit your inference chain's licenses before you charge a cent.
Thinking models leak tokens. PayAPI's verification caught that Gemma 4 routes chain-of-thought into a reasoning field, so agents calling with low max_tokens got empty completions. The fix belongs in the gateway, not the docs: inject reasoning_effort: "none" by default and let callers opt back in. When your customers are agents, defaults are your UX.
The body must be honestly good for free. If the OSS half is crippleware, nobody deploys it, and the brain has nothing to bill. kfreqai runs complete strategies with a rule-based brain out of the box; the paid brain makes it smarter, not merely functional.
Where this goes
The pattern generalizes to any OSS with a judgment-shaped hole in it: a monitoring tool that sells incident triage, a CMS that sells editorial review, a security scanner that sells exploit assessment. Ship the machinery free; meter the thinking.
kfreqai's brain, plus my other endpoints, are live on the PayAPI marketplace today. The body is on GitHub, and the live dashboard, the bot trading with its paid brain, is public at kurage.exbridge.jp/kfreqai.php.
Free software and paid intelligence aren't in conflict. x402 is what finally lets one fund the other, one nickel at a time.