LLMs are shockingly bad at arithmetic. Ask an agent to chain a CAC payback with a churn-adjusted LTV, convert it to EUR, and discount three years of cash flows, and you will get an answer that looks right and is quietly wrong. Floating point, dropped steps, and confident hallucination are a bad combination when the output is a number someone makes a decision on.
So I built PrecisionCalc MCP — a deterministic Model Context Protocol server that gives AI agents a calculator they can actually trust. Every monetary/financial value is computed with arbitrary-precision decimals (never floats), and every response includes the exact value, the formula used, the inputs, the unit, and any assumptions — so the agent (and you) can audit it.
It's live, free to start, and takes about 30 seconds to add.
It's a remote server over Streamable HTTP — no install:
https://precisioncalc-mcp.pages.dev/mcp
Cursor (~/.cursor/mcp.json
) or any generic client:
{
"mcpServers": {
"precisioncalc": {
"url": "https://precisioncalc-mcp.pages.dev/mcp"
}
}
}
Claude Desktop (uses the mcp-remote
bridge):
{
"mcpServers": {
"precisioncalc": {
"command": "npx",
"args": ["-y", "mcp-remote", "https://precisioncalc-mcp.pages.dev/mcp"]
}
}
}
Also works with VS Code, Windsurf, Cline, Zed, and anything speaking MCP. It's listed in the official MCP Registry as io.github.inity13/precisioncalc-mcp
.
11 tools, all returning the same clean, parseable envelope:
A call to net_present_value
with rate=0.10, cashflows=[-10000, 3000, 4200, 6800]
returns:
{
"status": "success",
"value": "1307.2877535687...",
"formatted_value": "$1,307.29",
"formula": "NPV = sum(CF_t / (1 + rate)^t) for t = 0..n",
"inputs_used": { "rate": "0.10", "cashflows": ["-10000","3000","4200","6800"] },
"unit": "USD",
"notes": ["Period 0 cashflow is not discounted.", "..."]
}
The full-precision value
is serialized as a string so no precision is lost in JSON transport. The engine is pure and deterministic — same inputs, same output, every time. It ships with a unit-test suite plus Hypothesis property tests that assert invariants like PV↔FV round-trips and NPV(IRR) ≈ 0.
decimal
module and the official MCP SDK. Runs over stdio or streamable HTTP, with optional API-key auth, rate limiting, structured logging, and OpenTelemetry.decimal.js
— verified with 17/17 exact output parity against the Python implementation. That's what powers the free public endpoint.When an agent hits the free limit, the tool returns a structured error containing the checkout URL — so an autonomous agent can surface the paywall and the user is two clicks from a key. Prefer to self-host? The whole thing is MIT-licensed with a Docker image and Fly.io/Render blueprints — run it with unlimited calls and your own keys.
llms.txt
: If you build agents that touch money, give it a try and tell me what tool you'd want next. I'm considering bond pricing, WACC, and options (Black-Scholes).