MCP server | SocialPipe

The same endpoints are available to an AI agent over the Model Context Protocol, so a client can call SocialPipe itself instead of you writing glue around REST. There is nothing to install: point the client at one URL and give it your API key.

https://api.socialpipe.dev/mcp

Connect

Your key goes in an Authorization header as a bearer token. It is the same sp_live_… key the REST API uses, from your dashboard.

Claude Code:

claude mcp add --transport http socialpipe https://api.socialpipe.dev/mcp \
  --header "Authorization: Bearer sp_live_YOUR_KEY"

Claude Desktop, Cursor and other config-file clients:

{
  "mcpServers": {
    "socialpipe": {
      "type": "http",
      "url": "https://api.socialpipe.dev/mcp",
      "headers": {
        "Authorization": "Bearer sp_live_YOUR_KEY"
      }
    }
  }
}

Restart the client afterwards. It should list three SocialPipe tools.

Your API key can spend credits. Treat the config file like any other file holding a credential. Do not commit it to a repository, and use a separate key per machine so you can revoke one without disturbing the others.

The tools

Tool	What it does	Cost
`list_capabilities`	What is callable, with each one's parameters and credit cost. Narrow by platform or search term	Free
`fetch`	Call any of those 261 capabilities by name, with its own parameters. Returns the platform's JSON	1–50 credits
`x_user_profile`	One X profile record, as X returns it	1 credit
`x_user_tweets`	One page of an X user's timeline	1 credit

Four tools rather than one per capability, on purpose. 765 near-identical tools returning 765 shapes of raw JSON is a far harder choice for an agent than four — and their definitions alone would be around 99,000 tokens of your agent's context before it did anything. So the agent asks what exists when it needs to, instead of being told everything up front.

The two x_ tools are shortcuts for the questions agents arrive with most, callable without a lookup first. Everything else goes through fetch. Cached reads cost nothing, and a call that fails, comes back empty or is declined by the platform is never charged — including a call refused for a misspelt name or a missing parameter.

The capabilities that return pages of other people — followers, following, reposters, commenters, search results, 504 of the 765 — are not reachable from any of these tools. An agent that can pull a follower graph in one call will; a bulk retrieval of third-party personal data should be something a developer chose to write. They are all available over REST.