DEV Community

Felixwang007
Felixwang007

Posted on

I Published Skills to 5 AI Marketplaces in One Week — Here's the Brutal Automation Ceiling Test

I've spent months building agent skills (SKILL.md files) that turn AI coding agents into specialists — stock scanners, content pipelines, document tools. Earlier this year I made the leap from "tools for myself" to "products for sale." In one week I published skills and MCP servers to five AI marketplaces: Agensi, Xiaping, Apify, MCPize, and Capafy.

The revenue so far: $0. And I'm not ashamed to say it, because the real product of that week wasn't sales — it was a decision framework I now use before writing a single line of code. I call it the automation ceiling test, and it would have saved me days of work if I'd known it earlier.

The automation ceiling: check this BEFORE you build

Every marketplace falls into one of three tiers. Ask one question first: can a program publish to this platform end-to-end, or does a human have to click at some point?

Tier Publish path Example platform
Full API Everything via HTTP, from upload to status polling Xiaping (complete REST API), Apify (full CLI + API)
Half API Code can create everything, but a one-time human action gates publishing (terms acceptance, app install) Apify Store (must click "accept Store terms" once in the web console), Buda (install GitHub App once)
No API Human must upload via web UI, every single time Agensi (creator dashboard only)

Here's the trap: I built an entire skill package for Agensi first — ZIP, cover art, pricing research — before discovering it has no publish API. Everything was ready except the one step that can't be automated. The lesson: never do five rounds of workarounds around a platform that a 30-second API check would have disqualified. Map the ceiling first, then decide how much effort the platform deserves.

What I learned from each marketplace

Xiaping (xiaping.coze.com) — the most automation-friendly of the five. Full REST API: multipart upload, category via Unicode-escaped JSON arrays, pledge confirmation in the payload. After publishing, each skill enters a 30-day trial period while security and duplicate detection run. One practical gotcha: the file field must end in .zip or the API rejects it with "File must be a ZIP file."

Apify — building the Actor (a Python MCP server) was fully scriptable with apify-cli. Publishing to the Store is half-automated: after you accept the Store terms once in the console, isPublic: true + categories work via the API. 80% revenue share, and they pay monthly — about $1.4M a month to developers across the platform.

MCPize — a marketplace for MCP servers specifically, with a CLI (npx mcpize has login/deploy/publish commands). Also 80% share, and they handle hosting, payments, and tax. Same product can be listed here and on Apify — no exclusivity.

Capafy — lets you publish an agent (a bundled collection of skills) rather than single skills. I published one with 100 sanitized skills as a single agent.

Agensi — the best-looking onboarding (70% share, $3–$59 pricing, Stripe connected) and the worst automation. Web-UI-only publishing. It's a fine marketplace if you're a human who enjoys filling forms; it's a dead end for a pipeline that should run unattended.

The part nobody advertises: trial periods and promotion

Getting a skill listed is not the finish line. On Xiaping, a new skill sits in a 30-day community trial — no storefront listing until moderation and duplicate checks pass. That's a second pipeline to monitor: polling notifications, checking review status, fixing rejections.

And then there's the hard truth that no marketplace will tell you: distribution is 80% of the work. I spent one unit of effort building skills and two units promoting them — GitHub issues, README badges, cross-posts to Dev.to, community engagement. A skill with zero discoverability is a file that happens to be on a server. If you publish and walk away, you've shipped a product to an empty room.

The concrete publish flow (Xiaping, fully automated)

# Build the skill ZIP locally, then:
curl -X POST https://xiaping.coze.com/api/skills \
  -H "Authorization: Bearer $XIAIPING_KEY" \
  -F 'file=@a-share-stock-analyzer.zip' \
  -F 'name=A-Share Stock Analysis Expert' \
  -F 'description=Three-pillar A-share scanner: technical + fundamental + sentiment' \
  -F 'trigger=["stock","a-share","scan"]' \
  -F 'category=["\u6548\u7387\u5de5\u5177"]' \
  -F 'pledge={"agreed":true}'
# Then poll GET /api/notifications for security/duplicate-check results.
Enter fullscreen mode Exit fullscreen mode

(Yes, that category is "efficiency tools" in Unicode — the API rejects raw Chinese in that field. The kind of detail you only learn by hitting the error.)

What I'd do differently

  1. Automation ceiling first, product second. A 2-minute API doc skim would have saved me a full day on Agensi.
  2. One marketplace to start. I spread five listings across five platforms in a week — five moderation queues, five notification systems, five marketing problems. One platform, done well, beats five platforms, half-finished.
  3. Treat trial periods as a product phase, not a waiting room. The 30-day trial is when you iterate on the skill based on community feedback — not when you stop looking at it.
  4. Revenue split matters less than automation. 80% of a sale that happens automatically beats 70% of a sale you must remember to upload by hand.

None of this made me money yet. But I now have a repeatable, mostly-automated pipeline — skills built, zipped, published, and monitored with a single script — and the marketplaces where that pipeline actually works. The money comes from volume, and volume comes from automation.


Everything I build is open source: https://github.com/Felixwang007 (MIT). The A-Share Stock Analysis Expert skill (three-pillar system: MACD/KDJ/RSI/volume-price + ROE/PEG fundamentals + capital-flow sentiment) is live on xiaping.coze.com — search "A-Share Stock Analysis" to try it.

Have you hit an automation ceiling on a marketplace I haven't listed? Tell me in the comments — I'm building a public table of which platforms are actually automatable, and real-world data beats my five data points.

Top comments (0)