What Neumann does
Neumann provides MCP tools for analyzing plans built from uncertain inputs, formulas, and connections between model nodes. Its main calculation is Monte Carlo simulation: the service samples each uncertain input repeatedly, evaluates the model for each trial, and summarizes the resulting range.
The supported planning problems include project budgets and costs, delivery schedules, capacity, revenue, and investment measures such as net present value, internal rate of return, and payback. Results can include means, percentiles, per-period values, minimums and maximums, target or budget probabilities, and a run identifier for later queries.
How it works
A model describes its uncertain inputs, their distributions, the formulas that combine them, and the links between nodes. Before spending a run, check_simulation_model can identify invalid distributions, broken formulas, disconnected nodes, and loops without a delay. estimate_simulation_size_and_cost reports the model dimensions, estimated memory, complexity, runtime, size tier, and account cost without executing the analysis.
Long-running operations return a run identifier and continue in the background. get_simulation_results can report progress, wait for completion, or retrieve selected nodes, edges, metrics, periods, percentiles, thresholds, and histograms from a completed run. Only the caller that started a run can read it.
rank_inputs_by_impact performs one-at-a-time sensitivity analysis. It varies one input while holding the others at their medians, then returns a ranked data set suitable for a tornado chart. summarize_outcome_scenarios provides a shorter worst, median, and best-case view, with configurable percentiles.
Setup and configuration
Connect an MCP client to the Neumann endpoint at https://neumann.kromatic.com/mcp. Account creation is available through POST https://neumann.kromatic.com/api/v1/accounts; authenticated requests use an API key as a Bearer credential. The account and credit tools can show authentication state, remaining anonymous runs, credit balance, reset timing, and recent credit transactions.
Anonymous callers receive three free runs per UTC day. Anonymous models are limited to 1,000 trials and 20 nodes. Account runs use credits: small simulations cost one credit and large simulations cost five; the listed price is currently $0.50 per credit. Validation, estimation, distribution conversion, result retrieval, account checks, and feedback do not consume runs or credits.
Tools and capabilities
- Run Monte Carlo simulations and retrieve output distributions across periods.
- Rank uncertain inputs by their effect on a selected outcome.
- Produce configurable worst, typical, and best-case scenarios.
- Validate model structure before execution.
- Estimate simulation size, memory, runtime, and credit tier.
- Convert ranges into supported distributions and distributions back into interpretable ranges.
- Check run progress and request targeted percentiles, thresholds, or histograms.
- Buy credits when an authenticated account has a saved payment card.
- Send feedback without including personal data, secrets, user numbers, or conversation transcripts.
Supported distribution conversions include normal, log-normal, uniform, triangular, beta, gamma, exponential, binomial, discrete variants, constant, and Bernoulli. Several distributions, including Poisson, categorical, Dirichlet, multivariate normal, and constant arrays, are not supported by the range conversion tool.
Limitations and notes
A simulation call can take longer than roughly 40 seconds and then requires repeated result checks. Sensitivity analysis can take minutes on larger models because it runs the model for each input and tested value. Anonymous results remain readable for 24 hours and only from the originating connection. Completed runs retain statistics for nodes and edges, but access is restricted to the caller that created them.
Scenario labels require care: “worst” means the lowest numeric value, so the cheapest result is the worst case when the outcome is a cost. Scenario input values are estimates at the requested percentile, not necessarily the exact inputs from one common trial. Inputs whose distributions produce lists are skipped during impact ranking.