How to automate social media posts with Claude

This is the whole setup, screen by screen, with nothing skipped: a brand-new PostMCP account, one social channel connected, the MCP server URL copied, and Claude wired up so that “post this to LinkedIn on Tuesday morning” is a sentence rather than a chore.

Budget five minutes. Most of that is waiting on OAuth screens you have clicked through a hundred times before.

Before you start

• A Google, GitHub or Apple account to sign in with. No card, no password to invent.
• One social account you are allowed to post from. LinkedIn is the easiest first one.
• Claude — either claude.ai on a paid plan (custom connectors are a paid feature) or Claude Desktop, which is free and reads a local config file instead.

Step 1 — Create the account

Open the sign-in page. There is no sign-up form to fill in and no email confirmation loop — the three buttons are the whole thing, and the same button signs you in next time.

You arrive in a workspace of your own, with the Accounts tab open and nothing connected yet. That empty state is the next thing to fix, because an assistant with no channels can draft but cannot ship.

Step 2 — Connect a social channel

Press Connect New Account at the top right. Seven networks are offered; each one hands off to that network’s own OAuth screen.

Choosing LinkedIn sends you to linkedin.com to approve the permissions. Come back and, if your credential covers several profiles or company pages, you pick which ones to sync — a personal profile and a company page are separate destinations, and only the ones you tick are ever posted to.

Step 3 — Copy your MCP server URL

Open AI Integration in the sidebar. This tab exists to hand your workspace to an AI client, and it knows about a dozen of them — Claude, ChatGPT, Claude Code, Cursor, Codex and the rest. Claude is the first card.

Selecting Claude reveals three steps. The first is the only one that matters here: your personal server URL, with your API key already embedded in the query string.

Step 4 — Add the connector in Claude

The Open Claude Connectors Settings button in step 2 above takes you straight there. Otherwise, in Claude open Settings → Connectors — the list shows what is already wired in, and the Add button sits in the top right.

1. 1
   Add → Add custom connector

   The Add menu offers two things. Browse connectors is Anthropic’s directory; Add custom connector is the one you want.

2. 2
   Name it

   Whatever you type here is what you will see in Claude’s tool list. “PostMCP AI” does the job.

3. 3
   Paste the server URL

   Into the Remote MCP server URL field, exactly as copied. There is no separate API key box — the key travels inside the URL — so there is nothing else to fill in. Press Continue.

Claude connects, reads the server’s tool list, and the publishing tools become available in your next conversation. On Claude Desktop the same thing happens through a config file instead:

{
  "mcpServers": {
    "postmcpai": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.postmcpai.com/mcp?apikey=pmcp_sec_•••"]
    }
  }
}

Quit Claude Desktop completely and reopen it afterwards. A window reload is not enough, because the server is started when the app launches.

Step 5 — Prove it works before you trust it

Do not open with a publish. Start with a question that cannot touch anyone’s feed:

“What social accounts do I have connected on PostMCP?”

Claude calls get_connected_accounts and reads back your profiles along with the profileId it will use to target each one. That single answer proves the whole chain at once — key valid, server reachable, backend up, channel connected — while nothing is at stake.

Step 6 — The first real post

Now describe the outcome and the constraints, and leave the wording to Claude. A prompt that works well in practice looks like this:

“Draft a LinkedIn post about the analytics dashboard we shipped today — first person, no hashtags, under 1,200 characters. Preflight it, show me the copy, and do not publish until I say so.”

Three things happen. Claude writes the copy. It calls preflight_post, which is a dry run that catches character-limit overflows, unconnected profiles, missing media and the credit cost without publishing anything. Then it stops and waits.

You read the draft, ask for a tighter second paragraph, and then say publish. That triggers create_post, and Claude comes back with the live URL. To schedule instead, name the time and the zone — “Tuesday 9am IST”, never a bare “Tuesday 9am”, because an unqualified hour is interpreted as UTC.

The same thing without leaving PostMCP

If you would rather not set up a client at all, the dashboard has the same agent built in. AI Agent in the sidebar opens a console pointed at the workspace you are already in.

Ask it for a draft and it does exactly what Claude would: reads your connected channels, writes the copy, and stops for approval rather than publishing.

Every run shows which tools it used and lets you open the raw MCP transcript. When an automation misbehaves, that transcript is the difference between debugging and guessing.

Where the posts actually live

Whatever writes them — Claude, the built-in agent, or you — the posts land in one queue. The content calendar is that queue with dates on it, and you can drag a scheduled post to a new slot.

The composer is worth a look even if you never write posts by hand, because it shows the exact MCP tool call it is about to make — server endpoint, tool name, arguments, target profile ids.

The four things that go wrong

1. 1
   The timezone was left implicit

   Scheduling defaults to UTC. “9am” for an India-based account means 14:30 local. Name the zone every time a wall-clock hour matters.

2. 2
   The client was reloaded, not restarted

   Claude Desktop starts its MCP servers at launch. After editing the config, quit the app fully and reopen it — a new window is not enough.

3. 3
   A token expired quietly

   Posts stop appearing with no obvious error. Ask “does anything need reconnecting?” — that runs get_account_health and names the dead connections.

4. 4
   Everything was approved on autopilot

   Approve reads permanently; keep the prompt on create_post and publish_post_now. The approval dialog is the last cheap place to catch a bad draft.

Beyond Claude

Nothing above is specific to Claude. The same URL turns ChatGPT, Cursor, Claude Code, Codex or any other MCP client into a publishing surface — the AI Integration tab has the exact steps for each. Claude Code, for instance, is a single command:

claude mcp add postmcpai -- npx -y mcp-remote https://mcp.postmcpai.com/mcp?apikey=pmcp_sec_•••

And nothing is specific to LinkedIn either. Once a second channel is connected, “post this to LinkedIn and X, but rewrite the X version to fit 280 characters” is one instruction, and the per-network variants are handled for you. If you want the full picture of what the protocol is doing underneath, start with what a social media MCP server is.