Skip to main content
The MCP server URL is discovered out-of-band. For Paperzilla, that usually means your client or user starts with the documented endpoint from the docs, app setup instructions, or docs indexed via llms.txt: https://paperzilla.ai/api/mcp This URL does not come from OpenAPI.

​
What happens after the client knows the URL

Once the client is pointed at the MCP endpoint and has a valid MCP API key, it can discover the callable surface from MCP itself. In practice, the client initializes the MCP connection first, then calls:
  1. tools/list
  2. Optionally prompts/list
tools/list returns the available tool names plus input schemas, so the client can see what calls exist and what arguments they take. That includes feed browsing and search tools such as:
  • feed_get
  • feed_search
  • paper_get
  • paper_markdown
That means MCP is self-describing after connection.

​
Does MCP have its own endpoints?

Externally, no per-tool endpoints are required. Paperzilla exposes one MCP endpoint at /api/mcp, and tool calls happen over that protocol connection. Internally, those tools are backed by Paperzilla’s own app logic and data access. A tool may read directly from internal services or data stores instead of calling the public REST API.

​
How this relates to OpenAPI

OpenAPI documents a REST API. MCP tool discovery works differently:
  • OpenAPI tells a REST client which HTTP routes exist
  • MCP tells an MCP client which tools exist through tools/list
So if you are building an MCP client for Paperzilla, you usually need:
  • The MCP server URL
  • An MCP API key
  • Standard MCP calls like initialize and tools/list
Use OpenAPI only for Paperzilla’s normal REST endpoints, not for MCP tool discovery. See also: