ai.mcpanalytics/analytics

Platform documentation and info: how it works, tiers, usage.

Direct link to the right account page for anything not doable in chat: billing, browser upload, report management. Hand the user the link and guide them.

Apply the user's layout wishes to the estimate's page through the layout agent; a new named arrangement, nothing overwritten, no number changes.

AI help desk: which analysis fits your question, interpreting results, fixing errors. Multi-turn.

A read of the data (average, count, total, highest/lowest by group, a value in a month) answered in this response, in seconds. Not a read -> immediate=false with the reason; continue with decide_path.

Ask a question across all your delivered analyses: a synthesized answer with citations back to specific reports.

Check a commissioned build in-chat: stage progress, queue position, rejection reason if the data didn't match the objective, honest ETA, report link when delivered.

Before naming a library tool: does it fit THIS dataset for THIS question? Column mapping, missing required inputs, method-fit verdict, the places it delivers. Read-only.

Commission a NEW analysis built for your question. tier is REQUIRED. The user picks. Easiest: fuzzy_request (plain language) + dataset_ref + tier. Snapshot = instant automated report (~2-10 min). JSON = a fast computed answer, numbers + method, re-runnable tool you own (~5 min). Brief = the computed answer on a one-page report: chart, numbers, method (~7 min). Deck = commissioned deep analysis, a durable re-runnable module you own (30-45 min). Failed builds are never billed.

List and search your uploaded datasets, with fuzzy matching on name, description, and tags. Returns each dataset's uuid:// reference for use in create_analysis and run_analysis.

Get your data in. Pass `data` as an array of row objects to create the dataset immediately and get a dataset_ref ready for create_analysis; omit it to get an upload link for a file only the user can reach. Add replace_ref (uuid://ID:KEY) with data to REFRESH an existing dataset in place; schedules and tools holding that reference read the new data on their next run.

Step 0 for a new question: which path answers it on this data. One record: route (reuse | answer | package | ask | none), a score with its reason for each of answer, package, ask and none, the compiled read plan when it is a read, the method family and the library's tool fit when it is a package, and the one question to ask when something is missing. Deterministic, read-only.

Browse the analyses you can run: the ones you commissioned plus the platform Standard Library (prebuilt tools; each result tagged source:'own' or 'standard_library'). Plain-language match; no query lists everything, your own first. Nothing fits? Commission it with create_analysis.

Before estimating: how did we answer this objective before, on this data or any data? Prior packages and library runs with their tools, mappings, bespoke module names, method and verdicts. Platform-wide, read-only.

Modify an EXISTING analysis into a new version: reword the question, swap the method, or add a variable. Pass tool_name + changes (plain language). Rebuilds on the analysis's own dataset by default; the original stays put. Returns pipeline tracking. Follow with build_status.

List and search the objects you own across every question: the curated charts, tables and figures of each delivered package, grouped by objective.

Order what the estimate promised after reviewing it: library tools that fit, a bespoke build, or both, computed on the whole dataset; one reviewed page delivered. Credits per tool run; failed runs never billed.

Read an analytics package back: status, every run under it, the report link once delivered.

Browse a delivered report's individual cards (charts, tables, insights) inline in chat.

Your report library: every analysis delivered, with status and links. Pass semantic_query to search report content in plain language.

Get a shareable browser link for a report, viewable without authentication.

START HERE for a new question: free, ~30 s. A rough answer over a sample plus the layout of the complete package, every place named with the question it will answer, and a page link. Then review_estimate with the user.

Run a delivered package again, on its own data or new data: the same tools, the same curated objects, the same layout, as a new package with its own link.

The estimate as you review it WITH the user: the question as understood, the estimated answer (sample, marked), every place and its question, the page link, a review checklist. Before order_analytics_package.

Run an analysis on your data. Returns a shareable interactive report URL with statistics you can cite, re-run and share, and the method named.

Standing re-runs of analyses you own: action='create' (weekly/monthly against a re-runnable data reference, connector:// or an https:// link; report emailed after each run), 'list', or 'cancel'.

Get an analysis's parameter schema. ALWAYS call before run_analysis.

Query your org's data warehouse free: browse the catalog (tables with column roles + computed metrics), semantically find data, plain-language ask, or named templates. Requires warehouse enablement (business plans).