Diceplots MCP private beta
A Model Context Protocol server that gives LLM agents a tool for exact dice math. Replaces "around 70%" hallucinations with 72.22%. Hosted today, licensable for self-hosting on request.
Why MCP
LLMs are routinely wrong about probability. Ask an agent "does 2d6+5 finish 11 HP?" and you'll often get a confidently wrong answer — sometimes within a few percent, sometimes off by 20+. The math is deterministic and computable in milliseconds; the only thing the LLM should be doing is formulating the question, not approximating the answer.
An MCP tool fixes that. The agent calls into the Diceplots engine for any dice question, gets back exact rationals, and explains the result in plain language. Same shape as a calculator tool: the LLM does the planning, the engine does the math.
What the tool exposes
-
strike(expression, hp)— full distribution + kill probability for a single strike. -
compare(expression_a, expression_b, hp)— side-by-side kill probabilities. The classic "which build kills faster" question. -
strikes_to_kill(expression, hp)— expected number of strikes plus the full cumulative kill curve. -
rounds_to_kill(damage, healing, hp, rounds)— Markov rounds-to-kill with optional per-round healing.
Every tool call returns exact rationals (numerator and denominator) plus a float approximation, so the agent can pick the right precision for its response.
Hosted vs licensed
- Hosted: point your MCP-aware agent at the Diceplots MCP endpoint. Per-call rate-limited; we run the engine and the math.
- Licensed self-hosting: run the MCP server yourself for air-gapped or compliance-bound deployments. Same Rust engine the website runs; same wire format as the hosted version so configs are portable.
Beta access
Hosted MCP is in private beta. Email with the agent runtime you'd plug it into (Claude Desktop, Cursor, custom MCP host) and the rough call volume — we send back a hosted endpoint and a key, typically same-day.