Agents · MCP

Drive workbooks from your agent.

The CLI ships an MCP (Model Context Protocol) server. Point any MCP-aware client at it and an assistant can list groups, scaffold projects, build, publish, and manage runtime settings as structured tool calls — no scraping, no glue scripts.

Claude Code

Either install the official skill (recommended — it ships with Workbooks-specific authoring guidance) or wire up MCP directly.

Option A — install the skill

curl -fsSL https://workbooks.sh/skill | sh

Restart Claude Code; workbook init, workbook publish, and authoring help are available via the /workbooks skill.

Option B — wire up MCP

Add to your .mcp.json (project or user-level):

{
  "mcpServers": {
    "workbooks": {
      "command": "workbook",
      "args": ["mcp", "serve"]
    }
  }
}

Cursor

Cursor reads .cursor/mcp.json with the same shape as Claude Code's config. Drop the snippet above into that file and restart Cursor.

Codex / other clients

Any MCP-compatible client works — Continue, Zed, Aider, custom LangChain agents. Stdio transport; no network ports to manage.

What the agent can do

• list_groups — your groups across the workspace
• list_workbooks — workbooks inside a group
• init_project — scaffold a new workbook project at a path
• build — compile a project to dist/
• publish — upload a built artifact to Studio
• inspect — print metadata for a workbook file
• unbundle — recover source from a built file
• get_runtime_config / set_runtime_config — manage WB_ENV values

Authentication

workbook publish uses the same OAuth flow whether invoked by you or by an agent. The agent triggers the flow once via a browser; the token is stored in your OS keychain. Agents cannot read the token directly — only the CLI can.