Marketing

How James Devonport Runs Ads, Blog and Social on OpenClaw Agents

Built by James Devonport (NoCode SaaS) · · Source: Substack

Quick answer

James Devonport moved a local Claude Code ad system onto OpenClaw on a $4 a month Hetzner server. Agents now report on Meta ads daily, write and publish blog posts, pull Google Analytics data and draft LinkedIn and X posts, with approvals by emoji in Slack.

Result: Devonport says they no longer log into Meta Ads Manager, ad reports arrive in Slack every day, and the social agent has been fairly reliable for about a week. They give no traffic, revenue or ad performance numbers.

What they built

James Devonport, who writes the NoCode SaaS newsletter, already had a Claude Code project that created and managed Meta ad campaigns. The catch was that it only ran on their laptop. They tried Claude Code's /loop feature to schedule it, but that needs a terminal open on a machine that is switched on, and they travel a lot.

So they moved the whole thing to OpenClaw on Hetzner's cheapest server, which they say costs $4 a month. It now runs on a schedule and messages them in Slack when it needs a decision. According to Devonport, there are four pieces:

  • Ads monitor: sends a daily report on how each Meta ad is doing, with recommendations.
  • Blog writing skill: a monthly job makes a content plan, then a post is written and published each day with illustrations generated by Gemini.
  • Web analytics agent: pulls data from the Google Analytics API. Devonport hopes to use it to improve ad messaging and aim blog posts at weak spots.
  • Social media agent: suggests LinkedIn and X posts each day based on blog posts, product updates and reviews, and makes the images with Gemini.

How it works

Devonport says setup was quick. They created the server, gave Claude the IP address and the official OpenClaw install docs, and let it run with permission prompts turned off on the clean VPS. They say OpenClaw was running about five minutes later. Then they described the local Claude Code setup and asked Claude to write a plan for moving it over.

They also built a private Slack app that posts to a #marketing channel and listens for replies. Reports and draft posts show up there, and approvals happen with emoji reactions. A checkmark on an ad report asks for three new ad variations. A checkmark on a social draft publishes it. They can also reply in the thread to ask for changes, and they say they never open the OpenClaw interface.

The article does not say which model OpenClaw runs on or what the model usage costs.

The result

Devonport says they no longer log into Meta Ads Manager and that the daily ad reports are working very well so far. The blog and social agents are newer. They say the social agent has been fairly reliable for about a week, though the writing style still needs work. They share no numbers on traffic, ad performance or savings, and mention they may release the code as a starter kit.

How to build something similar

  • Get one workflow working locally first, such as an ad report, before you put anything on a server and a schedule.
  • Pick one chat channel for everything the agent sends you, and make approval a single reaction or reply so it takes seconds.
  • Keep a human approval step on anything that spends money or posts publicly until you trust the output.
  • Only skip permission prompts on a fresh server with nothing else on it, and budget for model costs on top of hosting.

The original post

James Devonport · View on Substack

Want to build something like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week.

Join Shipping Skool — Free

More use cases

← All AI agent use cases