Guides

Coding agents

Let Claude Code or Codex create agents, automations and records in your app over MCP.

During development

Connect Claude Code or Codex to your running dev server:

pnpm dev
pnpm exec agentic plugin claude   # or: pnpm exec agentic plugin codex

agentic plugin writes a project-local plugin to .agentic/plugin (an Agentic skill and an MCP server) and registers it in the project's Claude Code or Codex settings. Run it once per checkout; .agentic/ is ignored by git.

The coding agent then gets the same tools as the API: create and update agents and automations, work with your Agentic Models, and inspect integrations and approvals. It acts as one of your dev accounts, with that account's role and approvals.

--user EMAIL or AGENTIC_MCP_USER picks the account when several exist.
--workspace ID or AGENTIC_MCP_WORKSPACE picks a team workspace instead of personal.

The bridge works only while the dev server listens on localhost, with a private capability readable only by you. It is never part of a build. Turn it off with agentic: { localMcp: false } (Nuxt) or voidAgentic({ localMcp: false }) (Void).

Local runtimes for your agents

To run an agent with your local CLI login during development, choose Codex CLI or Claude Code in the model picker. For a code agent, set runtime: codex or runtime: claude.

These runs disable the CLI's tools, plugins, web search, and shell access. Turn off local runtimes with agentic: { localRuntimes: false } (Nuxt) or voidAgentic({ localRuntimes: false }) (Void).

In production

Deployed apps serve MCP at /api/agentic/mcp with the same tools, access rules and approvals. It uses the dashboard's session, so it suits clients that sign in through the browser. Switch it off with the mcp feature toggle.

Features

Switch built-in features off, or package your own routes, services, tools and tables as a feature.

Tools

Give agents typed functions with defineAgentTool, and hold risky calls for approval.