Skip to main content

Installing Agent Skills

Rill's agent skills are bundled with the Rill CLI and generated directly into your project with rill init --agent. This page covers installation for each supported tool.

Requirements​

Install with the CLI​

New project​

Run rill init interactively and select an option at the "Agent instructions" prompt:

rill init
? Project name my-rill-project
? OLAP engine duckdb
? Agent instructions all

Or pass the --agent flag directly:

rill init my-project --agent all

The all option generates the skills in every supported format, so the project works with any agent out of the box. To generate files for a single tool only, pass claude, cursor, or agentsmd instead.

Existing project​

To add agent skills to an existing Rill project, run rill init with only the --agent flag:

rill init ./my-existing-project --agent all

The command only writes agent instruction files and MCP configuration; the rest of your project is left untouched.

Existing agent files are overwritten

If the project already has agent instruction files at the generated paths — for example .claude/CLAUDE.md, .cursor/rules/AGENTS.mdc, or an AGENTS.md at the project root — the command replaces them without prompting. Back up or commit any customizations first, and keep your own instructions in separate files so they survive regeneration (see Updating agent skills).

Options​

The --agent flag accepts the following values:

OptionDescription
claudeClaude Code skills in .claude/ (default)
cursorCursor rules in .cursor/
agentsmdTool-agnostic AGENTS.md format
allAll of the above
noneNo agent instructions
Commit the generated files

The generated files are plain text and designed to be committed to version control. Check them in so everyone on your team — and every agent — works from the same instructions.

Claude Code​

The all and claude options generate:

  • .claude/CLAUDE.md — entry point that tells Claude Code to load the Rill development skills before making changes
  • .claude/skills/rill-*/SKILL.md — one skill per topic: the development workflow, data analysis, and each resource type (see the skills reference)
  • .mcp.json — connects Claude Code to Rill Developer's local MCP server

Open Claude Code in the project directory. When prompted, trust the project's MCP configuration so Claude Code can connect to the Rill MCP server. The skills activate automatically based on your requests.

Cursor​

The all and cursor options generate:

  • .cursor/rules/AGENTS.mdc — an always-applied rule that points Cursor at the Rill development instructions
  • .cursor/rules/*.mdc — rules for the development workflow, data analysis, and each resource type, applied automatically based on their descriptions
  • .cursor/mcp.json — connects Cursor to Rill Developer's local MCP server

Open the project in Cursor and enable the rill-developer MCP server when prompted.

AGENTS.md-compatible agents​

This format works with agents that follow the AGENTS.md convention, including OpenAI Codex, Gemini CLI, and GitHub Copilot. The all and agentsmd options generate:

  • AGENTS.md — entry point at the project root
  • .agents/skills/rill-*/SKILL.md — one skill per topic, in the same format as the Claude Code skills
  • .mcp.json — MCP server configuration for agents that support it

The entry point instructs the agent to load the rill-development skill by name. Agents that support the Agent Skills convention resolve it from .agents/skills/ automatically. If your agent only reads AGENTS.md, the skills are plain Markdown files — point the agent at .agents/skills/rill-development/SKILL.md directly.

Connect the MCP server​

The generated MCP configuration points at Rill Developer's local MCP server:

{
  "mcpServers": {
    "rill-developer": {
      "type": "http",
      "url": "http://localhost:9009/mcp"
    }
  }
}

The server becomes available when you start Rill Developer:

rill start my-project

Through the MCP server, your agent can check resource status, inspect table schemas, run SQL and metrics queries, and read and write project files with immediate feedback on parse and reconcile errors.

Next steps​