@jarroba/mcp-calc

An MCP server for the maths a language model gets wrong — and, above all, for the ones it gets wrong silently.

claude mcp add jarroba-calc -- npx -y @jarroba/mcp-calc

// Claude Desktop, Cursor, Cline, VS Code…
{ "mcpServers": { "jarroba-calc": { "command": "npx", "args": ["-y", "@jarroba/mcp-calc"] } } }

5 tools · ~1.7k tokens of context · stdio, runs on your machine · nothing is uploaded, no API key.

Why this exists

A model that gets 17 * 23 wrong is not the problem: that one is easy to catch. The problem is the answer that looks right.

• 5 km + 300 m. Not a hard sum — but is the answer 305, 5.3, or 5300? A dimensional slip produces a number, not an error, and the number travels into the next formula.
• x^3 - 1 = 0. «x = 1» is true, and it is also only a third of the answer. A tool that stops there and says nothing teaches the caller that there was nothing else.
• 9.8 m/s^2 * t^2 / 2. Here m, s and t are all names of units and perfectly ordinary variable names. Read them wrong and you get a plausible quantity in the wrong dimension.

So every tool here reports how sure it is, not just what it found: exact or approximate, over which interval it searched, how many complex roots it is not showing, and which names it read as units. The one thing none of them will do is fill a gap with something plausible.

The tools

| | | |---|---| | calc_eval | Arithmetic with units and dimensional checking: 100 km/h to m/s, 5 km + 300 m, 70 kg to lbs, det([[1,2],[3,4]]). Adding metres to kilos is an error, not a silent number. Flags results that are not finite (NaN, ±∞), and says which names it read as units | | calc_solve | One equation (x^2 - 5x + 6 = 0, sin(x) = 0.5) or a linear system, one equation per line. Says whether the answer is exact or approximate, over which interval it searched, the degree of the polynomial and how many complex roots are not shown. Never invents a root it did not find | | calc_sheet | A multi-line calculation sheet with the result of every line. Variables carry downwards, units travel with the value, # starts a comment, ans is the previous result, and an error stays on its own line instead of killing the sheet. Declare currencies with # @moneda USD = 1.08 EUR — you supply the rate, so the sheet still computes the same thing next year. Optional Markdown table out | | calc_derive | Exact symbolic derivatives that keep the units: d/dx (3 m * x) is 3 m, not 3. Higher orders, and can evaluate the result at a point. Refuses to answer «0» when the variable is not in the expression, because a correct zero for the wrong question is the worst answer there is | | calc_steps | The worked steps, verified: a linear or quadratic equation, a derivative (naming the rule), or Gaussian elimination row by row. When it cannot explain a case it returns covered: false instead of writing a plausible derivation — which is the point, if a person is going to read it |

What it will not do

• No symbolic simplification. It was the obvious sixth tool and it was dropped after measuring: the engine leaves (x+1)^2 - (x-1)^2 untouched (it is 4x) and does not know sin²+cos²=1. A tool that hands back your own input under the label "simplified" is worse than no tool, because the caller cannot tell «already simplest» from «I could not do it».
• No integrals, no limits, no inequalities. Not supported, and said so here rather than discovered at the wrong moment.
• Steps beyond degree 2 for equations. calc_solve still answers exactly; what is missing is the explanation, and calc_steps says which cases it covers in every response.

Where it comes from

The engine is the same code that runs the calculator at jarroba.com/tools — pure TypeScript, ~200 unit tests, no server. This package is the whole catalogue filtered down to the calculation domain; if you want the other 95 tools (logs, tests, SQL, SPARQL, diagrams…), install @jarroba/mcp instead. You can also widen this one without switching packages: JARROBA_MCP_DOMAINS=calc,core.

MIT. Nothing here talks to the network. Source: github.com/Invarato/jarroba-mcp.