Simba documentation

Native result charts

Available in Simba MCP v0.16.0, published on PyPI. Your connected MCP server must provide v0.16.0 or a later compatible release, with a compatible backend and an MCP Apps-capable client. Package publication does not upgrade a hosted server: check its version and tool list. Native-client acceptance has not yet been observed and is recorded separately.

Three read-only tools display existing results in hosts that support MCP Apps:

Tool Inputs View
show_response_curves model_hash Response curves, served uncertainty and current spend where available
show_decomposition model_hash Contribution series in KPI units, with Overlap separately identified
show_optimizer_allocation model_hash, optional run_id Saved allocation spend and separate decision/comparison tables

Example prompts:

Replace these placeholders with authorised identifiers returned by list_models and list_runs. The tools do not create models, fit results or change budgets. Supplying a run ID selects that saved run; omitting it reads the model’s latest allocation result.

The view uses only returned values. Missing observations remain gaps. Legacy response grids with fewer than three finite points are disclosed rather than repaired or interpolated. Exact-value tables expose the supplied numbers. A current-spend marker is not an allocation recommendation. Response-curve reference levels and headline mROI can differ, so the view does not substitute one for the other.

Decomposition values are KPI units, not revenue. Overlap is a reconciliation term, not a channel. Allocation Revenue and ROI are decision quantities; OptimizedEvalRevenue, OptimizedEvalROI, HistoricalRevenue and HistoricalROI are accounting comparison quantities. These are presented in separate tables without computing uplift.

See the v0.16.0 tool reference for this version’s chart contract. Compare it with your connected server’s version and tool list before using the examples.

Compatibility and security

The server advertises ui://simba/charts.html with the MCP Apps MIME type text/html;profile=mcp-app. The packaged HTML contains inline SVG and script, requires no external resources and declares an empty resource CSP. No credentials enter the iframe. The existing structured and text results remain useful on clients without Apps rendering support.

The implementation follows the MCP Apps lifecycle and the shared MCP Apps UI protocol supported by ChatGPT. Documented protocol support is distinct from acceptance in a particular account or client version.

Automated fixture rendering covers the three views and missing-data conventions. Actual native-host acceptance has not yet been observed and is recorded separately. Do not describe an ordinary browser fixture screenshot as a live client screenshot.

Synthetic browser examples

These screenshots show the packaged view in a local synthetic test host at a measured 1280-pixel viewport. They are fixture evidence, not screenshots of native rendering in Claude or ChatGPT. Values are synthetic and do not represent a fitted customer model.

Response curves in the local synthetic test host

The response view keeps missing values as gaps and exposes the served numeric values.

Decomposition in the local synthetic test host

The decomposition view separates Overlap into a reconciliation panel rather than presenting it as a channel.