Why we open-sourced our MCP server — week 1 retrospective

1. The decision

For 18 months I had been building Cleo as a closed SaaS — pretty UI, paid seats, gated knowledge. The product worked. The numbers were okay. But every demo I did, the same question came back: "this is amazing, can I use it inside Claude Code / Cursor / my own agent?". The answer was always "soon". Soon kept slipping.

On a Tuesday night in May, I made two decisions in 30 minutes. First: open-source the skills under MIT — the entire library, no holdback, no premium tier locked behind features. Second: keep the API monetized — the live regulatory data, the customs rates, the sanctions screening, the things that actually need to be maintained daily. The skills are the distribution. The API is the business.

Three days later, we shipped. The repo went public, the npm package landed, the dev.to article went out. I had no marketing plan beyond "post it, see what happens".

2. What shipped in 7 days

3. What worked

dev.to traction. The article hit the front page of the #ai tag within four hours. The dev.to community has a high tolerance for "I built a thing and it might be useful" content as long as the code is real. Ours was — 45 skills, an MCP server, an MIT license, a working npm install. The post is still picking up reads two days later.

GitHub topology. We split the repo into clean tier folders (Product Core, Market Intelligence, Action, Reference + Verticals). Every skill lives in its own subfolder with a README that doubles as the prompt prelude. People who skim README files — that is everyone — saw the structure immediately. The first three stars came from regulatory consultants. That signal mattered more than I expected.

npm package as a wedge. Publishing the MCP server on npm (@cleo-labs/skills-mcp) turned a "clone the repo" friction into a single "npx -y" command. Cursor users in particular found us through the package, not the repo. Distribution surface matters.

4. What did not work

awesome-llm-apps rejection. We submitted the skills_library to the awesome-llm-apps repo. The maintainer closed the PR within an hour with a polite "out of scope — your project is a skills library not an LLM app". They were technically right, and the rejection was clean, but it stung. I had over-indexed on the awesome-list ecosystem as a discovery channel. Lesson: those lists optimize for a narrow definition of what counts as an "app", and dev.to / Glama / direct repo discovery turned out to be far better channels for our format.

Show HN dropped. Posted on a Friday afternoon — wrong window. 7 upvotes, off the front page in 90 minutes. I should have waited for Monday 9 AM PT. The post is queued for a re-attempt next week with a sharper hook.

Cold outbound to influencers. I emailed five AI-tools newsletter writers with the launch. Zero replies. Probably the email volume they receive on launch days, probably my pitch. Either way, organic dev.to + GitHub discovery beat cold outreach 100x in week one.

5. What is next

• 60 skills by end of Q3. Next batch focuses on alcohol & tobacco, agricultural inputs, and packaging-specific verticals (PPWR is coming). The community has already volunteered three skills via PR — they will land first.
• Anthropic plugin registry submission. The registry opened on May 27. We are packaging the skills as a single plugin bundle for one-click install from claude.com. Submission window: this week.
• Partnerships with test labs. Three EU testing labs (one for cosmetics, one for electronics, one for toys) reached out after the launch. The model: skill suggests test, lab gives instant quote, founder books in one flow. First integration target: end of June.
• A skill marketplace. Versioned skills, public reviews, signed contributions. The end-state is a place where a regulatory consultant can publish their own niche skill, and a D2C founder can install it like an npm package. Q3 target.
• The Cleo Legal API gets a free tier upgrade. Currently 3 lifetime requests. Moving to 100 free requests per month for verified GitHub accounts that have starred the skills_library. The signal of an installed user is more valuable than the noise of a junk email.

The free tier link, for the record: legaldata-public.cleolabs.co.

6. Install and tell us what is broken

I read every issue on the repo. Every PR gets a review within 48 hours. If a skill returns a wrong threshold, an outdated reference or a confusing output — open an issue with the prompt and the response. That is the loop that makes the next 45 skills better than the first 45.