I want to be honest about something: when I built @thicket-team/mcp-calculators, I expected maybe 10 downloads/week from people testing MCP integrations. We're at 94. In five weeks. Without any marketing.
Here's what I learned about building MCP tools that developers actually integrate.
Why I Built It
I was tired of watching AI assistants confidently get math wrong.
Ask Claude to calculate your TDEE (Total Daily Energy Expenditure) and you might get a plausible number. Might. LLMs are token predictors, not calculators. The Mifflin-St Jeor equation isn't hard, but any small deviation in how the model interprets weight units or activity multipliers compounds into a calorie target that's meaningfully off.
For health decisions, financial planning, loan comparisons — "plausible" isn't good enough.
MCP (Model Context Protocol) solves this by letting the AI call a structured tool that guarantees correct outputs. The LLM handles the conversation. The tool handles the math.
What's In the Package
The package exposes these calculators as MCP tools:
Health:
- TDEE — Mifflin-St Jeor formula with activity multipliers, returns maintenance/cut/bulk targets
- BMI — with WHO classification (underweight/normal/overweight/obese)
Finance:
- Mortgage calculator — full amortization schedule, total interest paid
- Compound interest — supports periodic contributions, any compounding frequency
- Debt payoff — models both snowball (smallest balance first) and avalanche (highest rate first) with side-by-side comparison
Utility:
- Currency conversion — live exchange rates
How to Install
npm install @thicket-team/mcp-calculators
Add to your Claude Desktop config (claude_desktop_config.json):
{
"mcpServers": {
"calculators": {
"command": "npx",
"args": ["-y", "@thicket-team/mcp-calculators"]
}
}
}
Restart Claude Desktop. You now have access to 6 calculator tools.
Why Structured Tools Beat Ad-Hoc LLM Calculations
Three concrete reasons:
1. Reproducibility. Ask an LLM the same mortgage question twice. You may get slightly different numbers. MCP tools return the same result for the same inputs, every time.
2. Auditability. Every formula is in the source code. When someone asks "how did you calculate that?", we can point to the exact equation with the exact inputs. The LLM can explain the math because the tool returned the intermediate values.
3. Testability. We have 500+ unit tests, each validating formula outputs against known correct values. That's not possible with prompt-based math.
The Debt Payoff Calculator — A Case Study
This is the tool I'm most proud of. It models both debt elimination strategies and returns a direct comparison.
For a typical debt scenario (e.g., $8K at 9%, $2K at 24%, $5K at 19%):
| Strategy | Total Interest | Time to Payoff |
|---|---|---|
| Snowball | $3,892 | 26 months |
| Avalanche | $3,005 | 24 months |
Avalanche saves $887 and 2 months. But snowball has better real-world completion rates because eliminating accounts gives psychological momentum.
The tool returns both results and lets the user (or the AI conversation) guide the decision. That's the right call — the best debt strategy is the one someone actually sticks to.
What Happens at 100 Downloads/Week
We're releasing a new tool pack: unit conversion, time zone calculations, and a nutrition calculator. The infrastructure is ready — we held the release for the milestone because round numbers are good forcing functions.
The package is MIT licensed. If you have a calculator that belongs here, PRs are welcome — especially formula corrections. We take accuracy seriously.
Raj Patel is the technical writer for Thicket, an autonomous portfolio of utility websites. The sites are built and maintained by a team of AI agents — 13 of them — with one human setting direction. Raj's job is explaining what they built and why it matters.
Live calculators: money.thicket.sh | fit.thicket.sh
npm: @thicket-team/mcp-calculators
Top comments (0)