AdeptAds has released a new set of agent skills and plugins designed to bridge the gap between AI coding agents and paid media management. The open-source repository, now live on Hacker News, provides the instruction sets and MCP connections necessary for Claude Code and Codex to autonomously research offers, plan ad groups, and draft creatives for Google Ads campaigns.

The Workflow: Human-in-the-Loop Launches

The core philosophy here is 'paused by default.' When an agent constructs a campaignβ€”say, a Search campaign for a plumbing service in Austin with a $40/day budgetβ€”it generates the assets and keywords but does not spend a dime until the user explicitly approves the launch. The agent can suggest up to 15 keywords, but will self-correct if suggestions don't fit the offer, as seen in the demo where it added 11 and rejected 4. This ensures that while the heavy lifting of keyword research and ad copy is automated, the financial trigger remains under human control.

Installation and Architecture

For Claude Code users, the setup involves adding the marketplace and installing the adeptads-google-ads plugin, followed by an OAuth sign-in via /mcp. Codex users follow a similar path with codex plugin add. For any other agent framework, users can install the standalone skill via npx and manually configure the remote MCP connection at https://api.adeptads.ai/mcp. It is crucial to note that the repository contains only the client-side instructions and plugin configurations; the actual campaign logic and MCP server run on AdeptAds' proprietary backend.

Cost and Limitations

While the plugins are free to install, the service requires an AdeptAds account with paid credits to run workflows. Google bills ad spend separately. The agents excel at building new Search, Display, and Performance Max campaigns, but they cannot handle autopilot setups, keyword-level bids, or negative keywords without manual intervention in the dashboard. Furthermore, while the agent reports progress in real-time during an active session, it does not push updates to closed conversations, meaning you must actively query it to check the status of queued tasks.

Key Takeaways

  • Campaigns are created paused: No budget is spent without explicit user approval to start.
  • Multi-framework support: Native plugins for Claude Code and Codex, plus MCP support for other agents.
  • Transparency: Agents provide a full list of changes and email alerts for significant financial actions.
  • Backend dependency: The logic runs on AdeptAds' servers, not locally, requiring an account and credits.

The Bottom Line

This is a smart play for the agentic era. By keeping the 'spend' button human-controlled, AdeptAds avoids the nightmare scenario of an AI burning through ad budget on a hallucinated keyword. It’s a practical, low-risk entry point for marketers who want AI speed but aren't ready for full autopilot.