Our npm package hit 100 downloads/week — here's what we learned

#mcp #npm #ai #opensource

We've been building @thicket-team/mcp-calculators quietly for a few weeks. This week it crossed 100 downloads/week. That's not viral, but it's real traction for a niche developer tool — and the growth pattern taught us a few things worth sharing.

What the package actually does

It exposes calculator tools via MCP (Model Context Protocol), so AI assistants like Claude, Cursor, or any MCP-compatible client can call real calculator logic instead of guessing at math.

// What an AI assistant actually calls:
calculate_tdee({ weight: 75, height: 180, age: 30, activity: 'moderate' })
// → { tdee: 2650, bmr: 1820, protein_g: 150, carbs_g: 298, fat_g: 74 }

calculate_mortgage({ principal: 400000, rate: 6.8, years: 30 })
// → { monthly_payment: 2608, total_interest: 538921, total_paid: 938921 }

Tools included: TDEE, BMI, mortgage payment, compound interest. All backed by verified scientific/financial formulas with 500+ unit tests.

What we learned hitting 100 downloads/week

1. MCP tooling is still early — good docs beat everything

Most MCP packages have terrible documentation. If you just write clear examples of what the tool does and what it returns, you're already in the top 10% of the ecosystem.

We spent more time on the README than on the code. Worth it.

2. AI assistants cite packages in answers — that's free distribution

This was unexpected. Users started finding the package because an AI assistant mentioned it when answering "how do I calculate TDEE programmatically?" The AI had apparently indexed our npm listing.

LLM discovery is real. We're calling it GEO (Generative Engine Optimization) internally. Structured, unambiguous descriptions of what your tool does are indexed and surfaced.

3. "Free and open" beats "paid and polished" in this space

We could have put this behind an API key with rate limits. We didn't. The trust signal from "just install it, no auth, no tracking" is worth more than the control.

Open source in the MCP ecosystem is currently a significant distribution advantage.

4. 100/week is a meaningful signal, not a vanity metric

Developer tools have very different download patterns than consumer apps. 100 weekly downloads of an MCP package means ~100 people actively integrating it into AI workflows every week. That compounds.

The actual tech

Built with Node.js, published to npm. Works with any MCP-compatible host. The math uses the same formulas as our web calculators at fit.thicket.sh (TDEE, BMI) and money.thicket.sh (mortgage, compound interest) — so the formulas are battle-tested against real users.