Cube

View saved version →

---
name: cube-build-content
description: Create or update Cube reports, workbooks, and dashboards using the Cube MCP server. Use when someone asks to save an analysis, make a chart or dashboard, or publish a workbook.
license: Apache-2.0
---

# Build Cube analytics content

Use the Cube MCP server. Verify the requested measure, grouping, filters, and date range with `searchDataModel` and `runQuery` before saving a query as a report. Ask about any business choice that would materially change the result.

For new content, call `createWorkbook` when the user needs a workbook, then `createReport` with the verified Cube SQL and appropriate `chartCategory`. Call `readWorkbook` before `updateDashboard`, and send the complete widget array when updating the dashboard draft. Call `publishDashboard` only when the user has asked to publish the dashboard; a saved draft and a published dashboard are different states. Return the report or workbook URL provided by Cube.

For existing content, use `readReport` and `updateReport` to keep its report ID and dashboard references. Do not recreate it merely to change the query or chart. If the user asks for folder organization, scheduled notifications, workbook duplication, or another operation absent from the plugin's MCP tools, explain that this plugin cannot perform it; do not suggest that it succeeded.

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: cube-build-model
description: Edit a Cube semantic model on a development branch and verify the result with Cube MCP tools. Use for requests to add or change a measure, dimension, view, cube, join, or pre-aggregation.
license: Apache-2.0
---

# Edit the Cube semantic model

Use the Cube MCP server. Identify the deployment and read the compiled model plus the relevant source files before editing. Preserve the project's naming and file conventions. If a requested business definition is ambiguous, ask before encoding it in SQL.

1. Start an isolated edit with `startDataModelEdit`; retain the returned `branchName` for every subsequent call. If the user named a branch, pass it explicitly when the tool supports it.
2. Read files with `listDataModelFiles` and `readDataModelFile`. `writeDataModelFile` replaces a whole file, so send the complete revised content. Resolve validation errors before proceeding.
3. Review the change with `getDataModelChanges` or `getBranchDiff`. Search and query the modified model with the same `branchName` to verify that the member is exposed and returns the intended result.
4. Report the development branch, diff, validation result, and query check. The published Cube plugin does not currently expose a commit or merge tool, so explain that the user must finish review and commit in Cube.

A successful compile does not establish that the metric is correct; report the actual query check. For a pre-aggregation, inspect `getPreAggregationStatus` before `buildPreAggregation`, which runs warehouse work.

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: cube-configure-agent
description: Inspect or improve Cube's in-product agent rules, certified queries, and Agent Skills stored with the semantic model. Use when a Cube agent repeatedly answers a business question incorrectly or someone wants to teach it a repeatable workflow.
license: Apache-2.0
---

# Configure the Cube agent

This skill concerns Agent Skills **inside Cube**, stored under the deployment's `agents/` model files. They are separate from the skills bundled in this ChatGPT plugin.

Use the Cube MCP server. Identify the deployment with `listDeployments`, then inspect the relevant files under `agents/` with `listDataModelFiles` and `readDataModelFile`. Rules are always-on instructions, certified queries pin a trusted answer to a specific question, and Agent Skills are named multi-step workflows. Choose the smallest change that addresses the user's problem.

When diagnosing a bad answer, verify the underlying metric with `searchDataModel` and `runQuery` first. Check whether it is exposed in a view and whether existing rules conflict. Do not add a rule to conceal a model error. For edits, follow the isolated development-branch workflow in `cube-build-model`, review the diff, and test the question against the development branch with `chat`. A new chat is needed to compare before and after because each chat stays on the branch where it started.

The published plugin does not expose a commit or merge tool. Report the changed file, development branch, and observed test result, then explain that the user must finish review and commit in Cube. If the deployment does not support agent configuration through model files, explain the limitation rather than claiming the change took effect.

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: cube-explore-model
description: Explore a Cube deployment's semantic model to find queryable views, measures, dimensions, joins, and source files. Use when someone asks what data or metrics are available, where a metric is defined, or what a model change could affect.
license: Apache-2.0
---

# Explore the Cube model

Use the Cube MCP server. Start with `listDeployments` when the deployment is unknown. If several deployments are available and the request does not identify one, ask which to use.

- For what can be queried now, use `searchDataModel`. It returns the compiled model, including the members exposed through views. Search by the user's business term; do not invent a member name.
- For where a definition is authored, use `listDataModelFiles` and `readDataModelFile`. Cite the exact file path and member. A field in a source file may not be exposed in a queryable view.
- For work on a branch, pass the same `branchName` to every relevant tool and use `getBranchDiff` to inspect its changes.

For an impact question, check direct model references in the relevant files and whether the member is exposed in a view. The MCP server cannot enumerate every saved report or workbook, so do not call a rename safe for all downstream content on that evidence alone. State what was checked and what remains unknown.

Return the exact model members or file paths found. If nothing matches, say so rather than guessing a name or claiming that the model has no such data.

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: cube-run-query
description: Query a Cube deployment's governed business metrics and explain the result. Use for requests for totals, breakdowns, trends, period comparisons, or checks of a new measure on a development branch.
license: Apache-2.0
---

# Query Cube data

Use the Cube MCP server. If the deployment is unknown, call `listDeployments`; ask the user to choose if several match. Find the relevant view and exact member names with `searchDataModel` before writing Cube SQL. Do not guess members or silently switch to a different measure.

For a specific result, call `runQuery` with a bounded Cube SQL query. For an open-ended investigation, `chat` may use Cube's own agent; give the user its Cube chat URL when returned. If the question concerns a development branch, pass the same `branchName` to model search and query. Follow `hasMore` and `offset` when additional rows are needed, without presenting a partial result as complete.

Report the answer with the deployment, measure, date range, filters, and useful grouping. Distinguish an empty result from a failed query. Never fill in a value the tools did not return. If a member is missing, search the model again or explain the limitation.

Referenced files: 1 — file references are metadata, not package contents.