Stop your agent from acting on overconfident guesses.
SimEngine gives an AI agent a real outcome-distribution calculator: schema-contracted Monte Carlo simulation, chance/CVaR safety constraints, and a proceed / escalate decision gate — over MCP, pure Python stdlib, zero dependencies.
Run it now
uvx simengine-mcpOr find it on the official MCP registry, or clone the engine — web UI, HTTP API, and CLI included. Python 3.8+ is the only requirement.
Why trust the numbers
The forecast pipeline is backtested on 2,819 walk-forward monthly forecasts over up to seven decades of public FRED data — every one produced through the real engine path and scored with proper scoring rules (Brier, log loss) plus calibration error (ECE) (full report):
| Task | n | Brier | ECE |
|---|---|---|---|
| CPI inflation > 0.25% MoM | 831 | 0.2317 | 0.081 |
| Industrial production falls | 1,168 | 0.2379 | 0.060 |
| Unemployment rate rises | 820 | 0.2515 | 0.171 |
This validates that engine-produced probabilities are calibrated, not that they beat markets. SimEngine models outcome risk — the spread of results given uncertain inputs — as decision support. It is not financial advice and issues no trade signals.
The decision gate
An agent hands decide_under_uncertainty a quick model of a decision
and the outcome range it can live with. The engine runs thousands of futures and
answers: “82% of simulated outcomes keep cash above zero — that clears your 80%
bar, proceed” — or escalate, with the tail risk spelled out.
Nine tools total: simulate, validate, optimize, Bayesian update, sensitivity,
convergence, baseline comparison, domain listing, and the gate.
Hosted endpoint — join the waitlist
A managed, agent-ready HTTPS endpoint is coming: no install, instant API key, free tier. Leave an email and you'll get an invite when it opens.
One email when the hosted tier opens. Nothing else.