# Skills and plugins

> Holistics skills and plugins for your own AI agent: which plugins are available, where to install them, and how to choose who writes AQL and visualizations.

The [MCP server](/docs/ai/mcp-server) gives your AI agent access to your Holistics workspace. Holistics skills teach the agent how to use that access well: how to explore datasets, write AQL, build visualizations, and develop AML. They're published as plugins in the [holistics/skills](https://github.com/holistics/skills) repository.

These plugins are for your own agent, such as Claude or Cursor. To extend Holistics' built-in AI instead, see [AI Skills](/docs/ai/skills).

## Available plugins

Each plugin bundles skills for one kind of work. Install the ones that match how you use Holistics.

| Plugin | Use it when |
| --- | --- |
| `holistics-development` | You develop your AMQL project with a coding agent such as Claude Code or Codex. See [Local agentic development](/docs/development/local-agentic-development). |
| `holistics-reporting` | You want to ask data questions on published datasets and dashboards with a general agent such as Claude.ai or ChatGPT. |
| `holistics-migrate-power-bi` | You're migrating a Power BI semantic model and reports to a Holistics AMQL project. Pair it with `holistics-development`. |
| `analytics` | You want generic analytics workflows (data exploration, analysis, reporting) that aren't tied to Holistics. |
| `analytics-finance` | You work with finance and SaaS metrics, such as lifetime value and unit economics. |

## Install

Install Holistics skills as a plugin whenever your agent supports plugins, since plugins also get versioning and updates. Each agent's setup page covers it:

- [Claude Web](/docs/ai/mcp-server/claude#step-4-install-holistics-skills)
- [Claude Desktop](/docs/ai/mcp-server/claude-desktop#step-4-install-holistics-skills)
- [Claude Code](/docs/ai/mcp-server/claude-code#step-4-install-holistics-skills)
- [ChatGPT](/docs/ai/mcp-server/chatgpt#step-4-install-holistics-skills)
- [Cursor](/docs/ai/mcp-server/cursor#step-4-install-holistics-skills)

For agents without plugin support, copy the skill files from the `plugins/<plugin-name>/skills/` folders in [holistics/skills](https://github.com/holistics/skills) into your agent's skills or instructions.

## Choose who writes AQL and visualizations

When your agent needs a query or a chart, it can either hand the task to Holistics AI or write it itself. The `holistics-development` and `holistics-reporting` plugins let you choose with the **AQL & Viz authoring** option.

### How each option works

- **`holistics-ai` (default):** your agent calls the `generate_aql` and `generate_viz` [MCP tools](/docs/ai/mcp-server/tool-list). These run Holistics AI sub-agents that write the AQL or visualization and return it to your agent. Like other Holistics AI usage, they're billed through your [AI models setup](/docs/ai/ai-billing#holistics-managed-models-vs-byok-vs-mcp): AI credits on Holistics-managed models, or your own provider on BYOK.
- **`self`:** your agent writes AQL and visualizations itself, using the AQL and visualization references bundled in the plugin. It then checks its work with tools that don't use Holistics AI, such as `validate_aql`. That usage is billed by your agent's provider (for example, your Claude plan) instead.

Tools that only read or run things, like `list_datasets`, `execute_aql`, or `search_dashboards`, don't use Holistics AI under either option.

### Set it up

This option requires `holistics-development` 0.9.0 or later, or `holistics-reporting` 0.2.0 or later. You can set it in Claude Code: see [Configure AQL & Viz authoring](/docs/ai/mcp-server/claude-code#configure-aql--viz-authoring).

### Trade-offs of having your agent write them

Writing AQL and visualizations in your agent has a few costs worth weighing:

- **Larger context window:** the AQL and visualization instructions stay in your agent's context for the whole session. Holistics AI sub-agents work in their own scoped context, so they don't add to your agent's context.
- **Less efficient prompt caching:** the `generate_aql` and `generate_viz` sub-agents are tuned to get the most out of prompt caching. Your agent might use more tokens to do the same work.