Give the model an engine instead of hoping.
Language models are confident and wrong about dates, business calendars and compounding. None of those live in a model’s weights. Hand it twelve tools over MCP and the arithmetic stops being a guess.
Shipped today. One command — npx -y @truecalc/mcp — gives any MCP client twelve tools over stdio. Nothing leaves your machine.
Edit any cell — it recalculates here, live
Worked examples
Two things you can do right now.
Five questions a model gets wrong
None of these are prompting problems. Each one has an answer that depends on an algorithm, a calendar or a rounding rule — none of which a language model has access to by thinking harder.
Open the demo →A workbook that survives the whole conversation
The MCP server holds a real workbook, so an agent can build a model over several turns, change one assumption, and recalculate — instead of restating the entire problem in every prompt.
Open the demo →Not quite your problem?
The same engine sits under all of these. The difference is only which surface you reach for.