July 4, 2026

Overview

AdLauncher is a white-label platform for Meta ads. A Singapore marketing agency uses it to run lead-gen campaigns for its clients. Staff launch campaigns from templates and track spend and leads for each client. An AI Copilot helps them write ads and find what is losing money. Clients sign in to a white-label portal. The product also runs an MCP server with 118 tools. Claude and Claude Code can use it to do the same jobs.

My role

Lead developer since February 2026, on a product the client owns. The client started it as a Lovable scaffold in January 2026, before I joined.

• 2,207 of the 3,933 commits (56%), the most of any contributor
• Lovable made 1,359 commits and the client made 367
• The client also codes on the product and started the Copilot and its Modes page

What I built

• AI Copilot: an agency chat agent that reads live ad data. I made most of its commits after the client started it.
• MCP server with 118 tools, plus the Connect Claude settings.
• Creative Studio: AI image and video creation with a prompt composer.
• Ads Manager table: status, budget, trend, spend and cost per lead for every campaign.
• Reports: health alerts for each ad account and billing status.
• Smart insights on the dashboard: what to scale and what is at risk.
• Campaign creation backend for the Meta Marketing API. Every new campaign starts paused.
• Marketing landing page and the saved audiences page.
• Tests: I made 122 of the 194 commits to the 105 test files.

The client started the saved audiences page, and Lovable scaffolded the Ads Manager table and campaign backend.

How I built it

I build with Claude Code agents. I write the plan and the task specs, and I set the checks each change must pass. I made 81 of the 89 commits on the repo's agent setup.

• A 533-line CLAUDE.md, the rule file the agents read, sets the rules.
• One change is one commit and one push. No agent may activate a real campaign.
• A tag on each task sets its checks, such as a plan review or a Playwright check.
• Up to six agents build at once, each in its own copy of the repo.
• For a large feature, one skill runs a team of 2 to 6 agents: builders, reviewers and a fixer.

The repo has 16 skills (reusable task instructions) and 4 subagents (specialist agents).

Tech stack

• Web: React 18, TypeScript, Vite, Tailwind CSS, shadcn/ui, TanStack Query
• Backend: Supabase (Postgres, Auth, 90 Edge Functions), Vercel
• Integrations: Meta Marketing API v22.0, Model Context Protocol, GPT Image 2, Higgsfield, Gemini
• Tests: Vitest, Playwright
• Agents: Claude Code, and Lovable until July 2026