Skip to main content
Cube agent skills are instruction packages that teach coding agents how to work with Cube. They can explore and change the semantic model, run queries, build dashboards, manage access, configure embedded analytics, and operate deployments. In Claude Code, the skills ship as the Cube plugin, bundled with the hosted Cube MCP server. In other coding agents, you install the skills on their own, then set up the Cube CLI and, optionally, the MCP server yourself.
Cube agent skills run in your coding agent. They are not Agent Skills in Cube, which your data team authors in the semantic model and runs from Analytics Chat.

​
Choose the right Cube integration

​
Install

​
Claude Code

Run these commands inside Claude Code:
Then run /mcp, select plugin:cube:cube, and sign in to Cube. Some skills also need the Cube CLI. The plugin uses the same endpoint as the Cube connector, so you don’t need both. For how Claude Code handles a server that is configured twice, see MCP scope precedence.

​
Codex, GitHub Copilot, Gemini CLI, and other compatible agents

Install the skills from skills.sh:
This installs the skills only. Install the Cube CLI below, and optionally add the Cube MCP server to your agent. The source is public under Apache 2.0 at cube-js/cube-agent-skills.

​
Install the Cube CLI

The skills use the Cube MCP server when it is connected, and the Cube CLI otherwise. The CLI is required for cube-admin, cube-embed, and most of cube-deploy, and for finding saved content with cube-explore-content — work the MCP server does not cover. On Linux or macOS, run:
For Windows installation and other options, see Cube CLI installation. Then authenticate with a browser:
For a headless environment, set CUBE_API_URL and CUBE_API_KEY instead. See Cube CLI authentication for both methods.

​
Available skills

​
Use the skills

Skills activate automatically when a request matches their description. You can also name a skill explicitly:
Exploration skills are read-only. Skills that write first inspect the current state and confirm the target deployment. Model edits happen on a dev-mode branch and reach production only when you ask to publish them — see How model edits stay safe for the one case where a commit can land on the deploy branch directly. Every MCP call and CLI command runs as the authenticated user, with their permissions and security context applied.

​
Troubleshooting

  • Cube tools are missing in Claude Code: Run /mcp, select plugin:cube:cube, and sign in.
  • The agent says the Cube CLI is missing: Install it from the Cube CLI reference, then start a new agent session.
  • CLI authentication fails: Run cube whoami. If needed, sign in again with cube login --url https://TENANT.cubecloud.dev.
  • The request targets the wrong tenant: Run cube context list and tell the agent which context to use.
  • A skill does not activate automatically: Name it in the request, for example, Use cube-explore-model to find the revenue measure.