Everything is supplied at call time — the arms with their pull and reward counts, the [asset][time] return matrix, the node and edge lists — so an answer is only as good as the data the caller assembles for it. The methods state their own limits: ARIMA wants at least 20 observations, simulate_montecarlo quantifies one random factor rather than a joint model, and solve_constraints will come back infeasible or unbounded instead of guessing.
A decision-intelligence server exposing optimisation, forecasting, risk and graph algorithms as tools, each taking its data as arguments on the call.
- Arm selection from trial history with UCB1, Thompson sampling or epsilon-greedy through optimize_bandit, and a LinUCB variant for situations where features change call to call through optimize_contextual.
- Constrained optimisation: solve_constraints for linear, mixed-integer and quadratic programs on the HiGHS solver, optimize_cmaes for continuous parameters in noisy gradient-free landscapes, optimize_evolve for discrete, integer and permutation spaces with an optional Pareto frontier.
- Task-to-slot assignment matched on energy level and duration with solve_schedule, and routing with plan_pathfind, which returns k-shortest alternatives as well as the best path.
- Forecasting and monitoring: predict_forecast with confidence bands, detect_anomaly with a Z-score or IQR test and the statistics behind the flag.
- Probability work: predict_bayesian for Beta-posterior belief updates, predict_ensemble for consensus across models, score_convergence for agreement between independent sources, score_calibration for Brier and log scores against observed outcomes.
- Uncertainty and sensitivity: simulate_montecarlo over a parametric distribution with percentiles and a histogram, simulate_scenario for what-if comparisons ranked by which variable moves the outcome most.
- Graph structure with analyze_graph — PageRank, Louvain communities, bottleneck nodes, an optional critical path — and portfolio downside with analyze_risk.
No account and no key. Some tools carry a [Premium] marker in their own descriptions: optimize_cmaes, solve_constraints, analyze_graph, analyze_risk, predict_forecast and detect_anomaly.
One command — npx -y @oraclaw/mcp-server