DEV Community: XM Yang

How indie makers in China actually get paid by overseas customers (with the real fee math)

XM Yang — Tue, 27 Oct 2026 07:35:07 +0000

If you build from China and sell to overseas customers, the product is the easy part. The hard part is the money plumbing: which channel actually lands money in a Chinese bank account, and how much of each sale survives the trip.

I spent weeks verifying the official fee pages and running the numbers. Here is the short version.

Route 1: PayPal direct, then wire to a Chinese bank

Best for direct client work. The catch is the withdrawal: PayPal China wires cost a flat $35 per withdrawal, minimum $150, 3–7 business days. That flat fee punishes small balances:

You withdraw	Flat fee eats
$150	23.3%
$350	10%
$700	5%
$1,000	3.5%

Rule of thumb: let the balance accumulate to at least $700 before wiring, unless cash flow forces your hand.

Route 2: Let a platform collect for you (Gumroad / Payhip)

Best for digital products — templates, ebooks, toolkits. The platform takes its cut per sale but handles cards, tax plumbing and payout.

What a sale is actually worth after the platform fee (card processing included, using each platform's published rates):

Price	Gumroad direct	Gumroad Discover	Payhip free plan
$10	$7.91	$7.00	$8.91
$29	$24.46	$20.30	$26.41
$99	$85.43	$69.30	$90.88

Two lessons: Discover/marketplace traffic is expensive (30% on Gumroad), and at small price points Payhip's 5% free plan beats Gumroad's 10% + $0.50 surprisingly often.

Route 3: Marketplaces pay you on their schedule

Etsy, app stores and freelance platforms (project fee 10% or $5 minimum on Freelancer, milestone escrow) remove the collection problem entirely — you trade margin and control for a built-in payout rail. Fine for validation, painful as a permanent home.

The stack I recommend

Client work → PayPal invoice, batch withdrawals at $700+
Digital products → Payhip or Gumroad as the cashier, drive your own traffic (never rely on Discover at 30%)
Validation → whatever marketplace already has your buyers

I turned the full research into a free bilingual playbook site — payment maps, platform thresholds, and 12 field-by-field playbooks (e-commerce, content, digital products, freelancing, POD, affiliate and more), each with a worked profit example and a first-week checklist: go.sweetai.work

If you're selling overseas from China, which channel burned you the most? I'd genuinely like to add the missing cases to the playbook.

"AI agent" is becoming meaningless — here's the test I use

XM Yang — Tue, 27 Oct 2026 06:30:37 +0000

Half the products calling themselves "AI agents" are chatbots with a new landing page. When I built the Personal Agents section on Tools Hub, I needed a definition that excludes them.

The test: does it do work in the background, on its own computer, with its own identity — while you're not in the chat? If it only replies when you type, it's a companion or a chatbot, not a personal agent.

By that test, seven products qualify right now: OpenAI Dots, Meta Muse, Grok Bot, Manus Cue, Gemini Spark, Claude Cowork and Perplexity Computer. Character.AI, Pi and Replika don't — they're excellent companions, but they wait for you. (Dot, the 2024 personal agent, shut down in October 2025 — a reminder of how fast this category moves.)

The full section, with what each one actually does: https://tools-hub.sweetai.work/best-ai-agents

Curious where you'd draw the line — is a browser agent that only acts while you watch an agent, or a macro with good PR?

I ranked AI models using only public benchmarks — and stopped publishing scores

XM Yang — Tue, 27 Oct 2026 06:29:00 +0000

Most "best AI" lists publish scores nobody can reproduce. Different agent harnesses, different prompts, vendor-reported numbers mixed with independent runs — the numbers look precise and mean almost nothing.

So on Tools Hub's rankings pages I made two rules. First: order only, no scores. Second: every board names its source on the page — SWE-bench Verified (independent runs compiled by Epoch AI) for coding, HealthBench for medical questions, LMArena's blind human votes plus EQ-Bench for writing, and Arena's text-to-image votes for image models.

The interesting part is where the boards disagree. EQ-Bench's judged board puts GPT first for creative writing; LMArena's human voters put Claude first. I follow the humans and say so on the page. That disagreement is more useful to a reader than a fake-precise 9.4/10.

Four boards are live (coding, medicine, writing, image). Happy to hear what other public benchmarks are actually worth ranking from — that's the bottleneck, not the page template.

https://tools-hub.sweetai.work/rankings