Is an MCP server just an API?

It is a program that speaks one specific protocol: JSON-RPC 2.0 messages over stdio or HTTP. The difference from a plain REST API is that the contract is model-facing — every tool ships a schema and a description the model reads — and uniform, so any MCP client can use any MCP server without integration work on either side.

Do I need to write code to use one?

No. Using a server means pasting a command or a small config block into your client, and each client documents where it goes. Writing your own server is code, with official SDKs in Python, TypeScript and other languages.

Is MCP only for Claude?

No. Anthropic released the protocol as an open standard, and it is spoken well beyond Anthropic's own applications — by Cursor and other editors, and by other model providers' tooling. A server written once works across them.

What is the current protocol version?

2025-11-25. Client and server negotiate a revision during the handshake, so a server typically also accepts older revisions from older clients.

LinkList — MCP

What is an MCP server?

An MCP server is a program that gives an AI application capabilities it does not have on its own — tools it can call and data it can read — over the Model Context Protocol, an open standard Anthropic released in November 2024.

Current protocol revision 2025-11-25 · facts on this page checked 2026-07-28

The short version

Claude, Cursor and other AI applications can only work with what is in the model's context. An MCP server feeds that context while the model works: it exposes named tools — search_sources, create_issue, run_query — and the model calls them mid-task, the way it already calls its built-in ones.

The application connects to any number of servers at once, and because the protocol is one open standard, any server works with any application that speaks it. Adding a capability stops being an integration project and becomes a line of configuration.

The three roles

Every MCP setup has the same three parts. The names come from the specification, and every client's documentation uses them.

Host: The AI application itself — Claude Code, Claude Desktop, Cursor. It runs the model and decides which servers to connect.
Client: The connector inside the host, one per server, holding one connection. Invisible in practice, but error messages name it, so the word is worth knowing.
Server: The program exposing the capabilities. It can run on your machine or behind a URL, and it knows nothing about which model is on the other end.

What a server exposes

The protocol defines three kinds of capability. Most servers in practice expose tools only.

Tools: Functions the model can call, each with a name, a JSON schema for its arguments, and a description. The model reads the description to decide when to call, so it steers behavior as much as the code does.
Resources: Data the host can read and attach as context: files, database rows, documents. The user or the application picks them; the model does not fetch them on its own.
Prompts: Reusable templates the user invokes deliberately — slash commands, in most clients.

Local and remote servers

Local (stdio)

The host starts the server as a process on your machine and talks to it over standard input and output. Right for anything that touches local state: files, a database on localhost, a browser.

Remote (streamable HTTP)

The server is an HTTP endpoint, reached with a URL and usually a key in the Authorization header. Nothing to install or keep running. https://linklist.io/mcp is one.

A server in use

A real session: Claude Code connected to LinkList's demo collection — an MCP endpoint over eight payroll-API documentation pages. The model calls one tool and answers with the source attached.

> what webhooks are available?

search_sources → 8 citations

[1] developer.tryfinch.com · fetched 2026-07-28 · link_id 3015

[2] plaid.com · fetched 2026-07-28 · link_id 3019

[3] docs.merge.dev · fetched 2026-07-28 · link_id 3021

The citations are the point of that particular server: each passage names the URL it came from and the date it was fetched, so the answer is checkable. Other servers return other things — files, query results, browser pages — but the shape is the same: the model calls a named tool and gets structured data back.

Try one in two minutes

The fastest way to understand an MCP server is to connect one and watch the tool calls. The demo collection publishes a live read-only key — no account required. Query the demo collection

Setup for each client, verified against the live endpoint:

Questions

Is an MCP server just an API?: It is a program that speaks one specific protocol: JSON-RPC 2.0 messages over stdio or HTTP. The difference from a plain REST API is that the contract is model-facing — every tool ships a schema and a description the model reads — and uniform, so any MCP client can use any MCP server without integration work on either side.
Do I need to write code to use one?: No. Using a server means pasting a command or a small config block into your client, and each client documents where it goes. Writing your own server is code, with official SDKs in Python, TypeScript and other languages.
Is MCP only for Claude?: No. Anthropic released the protocol as an open standard, and it is spoken well beyond Anthropic's own applications — by Cursor and other editors, and by other model providers' tooling. A server written once works across them.
What is the current protocol version?: 2025-11-25. Client and server negotiate a revision during the handshake, so a server typically also accepts older revisions from older clients.

MCP starter pack

Verified setup configs for every client.

Copy-paste configuration for Claude Code, Claude Desktop and Cursor, plus the eval harness so you can measure your own retrieval.

One email with the pack. No sequence.

Point your agent at sources you trust.

LinkList is a remote MCP server: your agent searches the collection you curated and answers with the URL and fetch date attached. Free tier, no card.