AI Assistants (MCP)

Every constrained parameter advertises its full vocabulary in the tool schema, derived from the enum that defines it rather than retyped: all 31 optimization methods, all 14 Indian and 3 US benchmarks, the return processes, stress scenarios, goal bases, drift modes, rebalance cadences, risk-free tenors, chart names and search operators. A client that never reads an MCP resource can still enumerate every legal value, and a single request may carry the entire method catalogue.

Common aliases such as “hierarchical risk parity” and “mean variance” normalize server-side, and an unrecognized method returns the offending names alongside the full valid list rather than a bare validation error.

Seven resources back this up: config://methods, config://method_requirements (the minimum assets each method needs, for example 4 for HERC2, stated before a model chooses one), config://benchmarks, config://us_benchmarks, config://constraint_capability (what each method does with each kind of constraint), config://constraint_presets (the SEBI presets a constraint block can name) and config://fund_manager_mode (the rules of Fund Manager Mode and what it does not add). Each is generated from the enum or table that defines it, so a new member cannot go missing, and each carries a description per entry.

Deliberately outside the MCP surface: authentication, billing, OAuth client management, account settings, the in-app chat endpoints, and the frontend's own UI state. These are session, security or presentation concerns rather than analysis capability.

Chart data is returned as data-only JSON packets under a 25 KB budget and the client renders them. Nothing is rendered server-side as an image, so an assistant can restyle, re-scale or combine a chart without a round trip, and no rendering assumption is baked into the payload.

get_monte_carlo_chart_data serves eleven packets: wealth_fan, goal_probability, terminal_quantiles, path_metric, drawdown_state, scenarios, metric_suite, plus the walk-forward set weight_paths, goal_ladder, annualized_return and paired_comparison. The last two take a view argument to select between series that do not share an axis.

The metric_suite packet plots the stored P5, P25, P50, P75 and P95 across paths plus the cross-path mean, and carries defined-path and observation counts. A missing risk-free benchmark, and the mathematically inapplicable simulated-path Deflated Sharpe and Sharpe Stability cases, fail with the engine's explicit reason rather than returning a zero benchmark or a misleading statistic.

Every paired_comparison view states in its footnote that the two arms ran identical paths and that a ratio above 1 is not evidence that re-optimizing outperforms. Monte Carlo results must always be presented with their assumptions block: probabilities are conditional on the model and the frozen history, and are never forecasts. See rolling-origin validation.