🚀 Lovable Agent Skills: Claude Code + Codex + More

Edit Lovable Projects 10x Faster—Without Leaving Your IDE

Stop copy-pasting between Lovable and Claude. Stop wrestling with two-way sync. Stop burning through API credits for simple changes.

The Lovable Agent Skills plugin brings the power of your coding agent directly to your Lovable projects. Codex, Claude Code, and other Agent Skills runtimes share the same provider-neutral workflows; Claude Code commands and hooks remain available for existing projects.

What You Get:

✨ Edit Lovable projects right in your IDE with Codex, Claude Code, or another coding agent 🧪 Preview Testing - automated test plans run against your live Preview app 🏗️ Dual-architecture support - works with both Vite SPA and TanStack Start (SSR) projects (v1.8.1) 🔌 Lovable MCP integration - deploy via API instead of browser (NEW in v1.8.0!) 🗺️ Project Structure Map - agents navigate your codebase faster ⚡ Auto-push to GitHub - automatic commit and push after every task 🚀 Auto-deploy to Lovable - no manual commands needed 🤖 Complete workflow automation - from code changes to production 💰 Save credits by using your Claude subscription for heavy lifting 🔄 Zero friction syncing - between your code, GitHub, and Lovable Cloud 🛡️ Automatic secret detection - never forget a required API key ⚡ Instant setup - generates project context automatically ✅ Verification built-in - tests run after every deployment

Provider-neutral agent contract

Version 2.0.1 includes the init skill and a shared YOLO authorization policy. Reinitialize an existing project to refresh its generated instructions without resetting its settings. The Codex manifest declares the official Lovable MCP server inline; .mcp.json retains the same direct server map for compatibility. Both configurations are checked for consistency by the tests.

Use the init skill ($lovable:init in Codex) to initialize or reinitialize a project. The Claude /lovable:init-lovable command uses the same workflow. Existing settings and custom instructions are preserved. deploy.yolo_mode: true grants standing authorization for in-scope Lovable prompts via MCP or browser; generic confirmation rules apply when YOLO is off. Explicit task restrictions and destructive-operation confirmation still apply.

New projects use .lovable-agent/ as the shared source of truth:

.lovable-agent/
├── config.json             # project, sync, deploy, testing, and secret metadata
├── context.md              # provider-neutral project context (no secret values)
├── preview-token.local     # ignored Preview credential, never committed
└── tests/                  # plans, profiles, and results

AGENTS.md is the canonical instruction file. CLAUDE.md and the legacy .claude/lovable-claude/test/ workspace remain compatible aliases. Run /lovable:migrate-agent