Postapop: An AI Social Media Manager With Its Own MCP Server
What It Is
Postapop is an AI social media manager for small business owners who are also their own marketing department. Its assistant, Rae, reads the page you share, learns your voice, drafts a content plan and writes the posts. You review them, schedule them, and Postapop publishes them on Instagram, Facebook, Threads, YouTube, Pinterest, TikTok, LinkedIn and Bluesky. I designed, built and run it, and it has paying customers.
What Rae Does
- Learns the brand from one page. Share a URL or describe the business in a sentence. Rae builds a brand card you can correct: goals, voice, colours, fonts and how often to post.
- Plans before it writes. A draft strategy with content pillars, then post ideas from your own saved articles, videos and voice notes, plus a Radar that follows chosen blogs, newsletters, YouTube channels and podcasts.
- Writes for each format. Captions for every channel, hook-and-beats scripts for TikTok and YouTube, and the text for each box in your own Canva designs, sized to fit.
- Schedules and reports. A calendar with suggested days and times, publishing on schedule, and every channel's numbers in one place with a Monday summary of what worked.
- Never posts on its own. Rae proposes and you decide. Nothing goes out until you schedule it.
Postapop Inside Claude
Postapop has its own remote MCP server, so customers can run their social media from inside Claude. They connect with OAuth and sign in with their Postapop account, Google sign-in included, and Claude can then read their brand profile, list their channels, suggest posts, draft from saved sources, find the next free slot in the calendar and schedule posts.
The same rule holds there as in the app. Claude confirms the text, channels and time with the person before it schedules anything, and asks again before publishing straight away, because that cannot be undone.
What Building the MCP Server Taught Me
- Keep the product rules in the API. The MCP layer checks inputs and shapes results. The web app and the AI connection then can't disagree about what is allowed.
- Errors should be sentences a model can act on. "Cut at least 76 characters, or leave Bluesky out of this draft" gets fixed on the next call. A stack trace doesn't.
- Separate read tools from action tools. Reading is safe to approve once. Every action does one small thing and says what it never does.
- Label text that other people wrote. Saved articles and newsletters reach the model marked as outside material, through read-only tools, so words inside them can't steer it.
I turned these lessons into an open-source course, the MCP server lab. If you would rather learn in a room with help, see my AI workshops.
Why It Matters for Client Work
Postapop is the same pattern I build for companies: an AI that works inside the tools people already use, does the drafting and digging, and leaves the decisions to a human. If you want your own product or internal tools connected to Claude, that's a project I take on.