MCP server setup

The MCP server is how an assistant reads and drafts inside the brands you grant. It is not a browser session. The steps below match the server that ships in the product.

Create a key

1. Sign in as an owner or admin.
2. Open Settings.
3. Name the client, check the brands, and review the grants. Read and Draft start on. Schedule, Publish, Setup, Automation, and Approve start off. See AI features.
4. Create the client and copy the key. It is shown once. The settings screen also shows the endpoint for the server you are running.

The key is stored as a hash. Revoking it stops the client on the next call. Putting the key in a query string is rejected. Do not commit the key or paste it into a ticket.

A local server uses http://127.0.0.1:8788/mcp. A hosted server prints its own origin. Use the address on the screen, not a guessed host.

Claude Desktop

In Claude Desktop, open Settings, then Developer, then Edit Config. Add a remote server. Replace the address and the key with the values from Settings.

{
  "mcpServers": {
    "automagic-social": {
      "type": "http",
      "url": "https://your-host.example/mcp",
      "headers": {
        "Authorization": "Bearer PASTE_THE_KEY_HERE"
      }
    }
  }
}

Restart Claude Desktop. The key belongs in the header, not in the address.

Cursor

Add a remote MCP server with the same address and header:

{
  "mcpServers": {
    "automagic-social": {
      "url": "https://your-host.example/mcp",
      "headers": {
        "Authorization": "Bearer PASTE_THE_KEY_HERE"
      }
    }
  }
}

Any other client

POST /mcp speaks MCP over HTTPS.

• Send Authorization: Bearer <key>.
• Send Content-Type: application/json and Accept: application/json, text/event-stream.
• Current clients use protocol 2026-07-28 and send one JSON-RPC message per request, with MCP-Protocol-Version, Mcp-Method, and the same version inside _meta.
• Older clients can initialize and then send Mcp-Session-Id.
• GET /.well-known/oauth-protected-resource/mcp describes the server for clients that use OAuth instead of a pasted key. A person signs in and checks the brands and capabilities. The assistant never handles that password.

tools/list only includes tools the token can call. A tool the client is not allowed to use is rejected before it runs.

What the tools do

Setup. List and create brands, rename a granted brand, read and update presets, list accounts, start a connection, check whether a connection needs a person to sign in again, and report missing developer-portal steps.

Content. Upload media from a URL or in chunks, create campaigns and drafts, save a version for each account, check it against that network’s limits, read TikTok and LinkedIn options, fill in the required choices, and submit the draft for review.

Scheduling and publishing. Schedule, reschedule, and cancel. Publish only when this client has Publish and the post is already approved. Read status and the home screen’s next actions. Reconcile an uncertain post without sending it again. Retry a failed post by turning it back into a draft.

Automation. Create, list, pause, resume, and delete a queue. The worker creates a normal draft. It does not approve it.

A prompt named plan_a_week asks the assistant to draft a week of posts and leave them for approval. Resources named automagic://brand/<id>/settings and automagic://brand/<id>/attention expose that brand’s presets and the home screen’s next actions.

Every tool calls the same services as the screens, including validation, brand limits, approval, jobs, and the audit log. Results and logs do not contain network tokens or the assistant’s key.