How to Let Claude Post to Social Media for You (MCP)

You ship all day, and the posting never happens. This guide is for founders who already use Claude Code (or Cursor, or any MCP client) and want the AI they code with to handle social media too. By the end, "just shipped dark mode, post about it" will produce a real, scheduled post.

What do you need before Claude can post for you?#

Three things: a scheduler that runs an MCP server, your social accounts connected to it, and an API key for your agent. Claude alone cannot post anywhere.

Claude has no access to X or LinkedIn by itself. MCP (Model Context Protocol) is the open standard that fixes this: a server exposes tools, and any MCP-capable agent can call them. Anthropic released it in November 2024, and by now every serious coding agent speaks it.

posthell runs a remote MCP server for social media. You start a trial, connect your networks through their normal logins (no X API keys of your own), and create an agent key in the dashboard's Agents tab. The key is shown once and can be revoked instantly.

How do you connect Claude to the MCP server?#

One command in your terminal. Claude Code registers the server and discovers the posting tools by itself; there is no plugin to install or code to write.

claude mcp add --transport http posthell https://www.posthell.com/api/mcp --header "Authorization: Bearer YOUR_API_KEY"

That is the whole integration. Cursor and VS Code take a small JSON config instead, and Claude Desktop connects through the mcp-remote bridge; the MCP setup page has copy-paste snippets for seven clients.

What happens when you tell Claude you shipped something?#

Claude calls two tools: one shapes your rough note into post angles in your voice, one queues the best draft with a suggested time. You approve it from your queue.

Here is the actual flow, tool by tool:

Tool	What it does	Needs your approval?
shape_post	Turns rough notes into 2-3 post angles, grounded only in what you wrote	n/a (drafting)
create_draft	Queues a draft with a proposed time	Yes, always
publish_post	Publishes for real, now or scheduled	Only if you opt the key in
list_posts	Reads your queue so it never drafts duplicates	n/a (read)
get_growth	Reads follower growth per network	n/a (read)

The drafts land in a "from your agent" section of your posting queue. You read the post, edit a word if you want, tap approve, and posthell publishes it to every network you selected, adapted per network so the LinkedIn version does not read like a tweet.

Can Claude publish without asking?#

Yes, but only if you flip that key to autonomous publishing. It is off by default, and X spending is hard-capped by your credit wallet either way.

The default is deliberate: agents draft, you approve. If you want true autopilot, you enable "posts without asking" on a specific key, and Claude then publishes to all your connected networks on its own, X included. The cap matters because X's API charges per post; the wallet means a runaway loop fails safe instead of running a bill. The automate pages list exactly what an agent can do on each of the 15 networks.

What does it cost?#

A posthell plan: Solo at $12/mo or Pro at $29/mo, both with a 7-day trial. MCP access is included on every plan, and only X posting is metered.

Solo includes 200 X posts a month and 40 AI drafts; Pro includes 600 and 150. Every other network (LinkedIn, Threads, Bluesky, Instagram and the rest) is unlimited. There is also a CLI, npx posthell-cli, that wraps the same tools for scripts and agents without an MCP client.

What we learned building this#

I built posthell's MCP server, and the first thing I did after wiring up autonomous publishing was turn it off by default. Watching my own agent work made the reason obvious: the drafts were good, but roughly one in three needed a human touch, a number changed, a sentence cut, a joke that did not land. Approve-with-edit takes ten seconds; a bad post is public forever.

The rate limits came from the same testing: 6 shape calls and 12 drafts a minute, 30 drafts a day. A daily posting workflow never notices them, and one misbehaving loop cannot burn a month of AI quota before breakfast.

Where to start#

Run the trial, connect two networks, add the MCP server with the one-liner above, and tell Claude one true thing you shipped today. If the draft that shows up in your queue is not something you would post, delete it and lose nothing. In my experience, you will approve it. If you are still deciding whether an agent should touch your accounts at all, read can an AI agent run your social media first.