What Is an MCP Server?

MCP server meaning, in plain English

An MCP server sits between an AI app and one product. The product might be Shopify, HubSpot, GitHub or a folder on your laptop. The server tells the AI app which actions exist, runs one when asked and hands back the result.

The MCP docs define servers as "programs that expose specific capabilities to AI applications through standardized protocol interfaces" (server concepts).

"Server" here means software. A local MCP server is a small process on your own computer. A remote one runs in the cloud like any web service. You do not buy hardware for it.

It is one of three roles in the protocol (architecture overview). The host is the AI app you use, such as Claude Desktop or VS Code. The host creates one client for each server you add. The server is the program at the other end. You only ever choose the host and the server.

For the protocol itself and why it exists, read What is MCP?.

What does an MCP server do?

A server can offer three kinds of things (server concepts):

Tools are the part that lets an AI act in your account. A tool is a name, a description and an input schema. The model reads the description and decides when to call it.

Here is the shape of a tool definition. The field names come from the spec. The tool itself is made up.

{
  "name": "find_invoices",
  "description": "Find invoices by customer and status",
  "inputSchema": {
    "type": "object",
    "properties": {
      "customer": { "type": "string", "description": "Customer name or ID" },
      "status": { "type": "string", "enum": ["open", "paid", "overdue"] }
    },
    "required": ["customer"]
  }
}

When the model calls a tool, the server checks the inputs against that schema, runs the action and returns text or structured data. For a SaaS product, running the action usually means calling the product's API with a credential the server holds.

How an MCP server handles a request

The model never reaches the product directly. It sees a tool list and tool results. The server holds the credential and decides which actions exist. That makes the choice of server the main control over what an AI can touch in your account.

The messages are JSON-RPC 2.0, and the current spec revision is 2026-07-28 (spec). You do not need either fact to use a server. If you are building one, Model Context Protocol: how it works has the wire-level detail, and MCP vs API covers how a server relates to the API underneath it.

Types of MCP servers

The spec has no list of server types. Four differences matter when you pick one.

Local or remote

A local server runs on your computer and talks to the AI app over stdio. A remote server runs at a URL and uses Streamable HTTP. Those are the two standard transports in the spec (transports).

Web-based clients need the remote kind. ChatGPT's developer mode connects to remote servers (OpenAI). Claude reaches custom connectors from Anthropic's cloud, not from your device (Claude help center). Remote vs local MCP servers compares the two in full.

Who builds and runs it

• Vendor servers come from the company behind the product. GitHub runs one, hosted by GitHub or run locally (GitHub). Sentry runs a remote one (Sentry).
• Reference servers are maintained by the MCP project to show how the protocol works. Filesystem and Git are two. The repo describes them as educational examples, not production-ready software (servers repo).
• Community servers are built by anyone. Check who maintains one, and when it was last updated, before you install it.
• Hosted platforms run servers for many products behind one account. PopMCP is one.

Read-only or read-write

A read-only server can look things up. A read-write server can also create, update or delete.

A tool can label itself with hints such as readOnlyHint. The spec tells clients to treat those hints as untrusted unless they trust the server (tools spec). So read the tool list yourself.

A few tools or hundreds

Some servers expose a handful of tools and others expose hundreds. A longer list covers more jobs, but every tool definition the client loads takes room in the model's context.

The MCP docs point clients that connect to many servers toward progressive tool discovery, which means not loading every tool up front (architecture overview). MCP token limits covers that trade-off.

MCP server examples

Servers you can use today

Older lists need care. Anthropic's launch post in November 2024 named pre-built servers for Google Drive, Slack, GitHub, Git, Postgres and Puppeteer (Anthropic). The reference versions of GitHub, Google Drive, PostgreSQL, Puppeteer and Slack have since been archived by the MCP project (servers repo), so check a server's README before you rely on it.

What it looks like on business tools

These examples use real tool names from PopMCP's catalog.

A Shopify store. You ask, "Which products had the most orders last week?" The model can call shopify_get_orders for the date range and shopify_get_products for the names, then build the table.

A HubSpot pipeline. You ask, "List open deals over $10,000 with no activity in 14 days." The model can call hubspot_deals_search with those filters. If you then ask for a follow-up task on each deal, it calls hubspot_tasks_create. That second step is a write. It needs a paid plan, and it runs as soon as the model calls it.

Three systems in one question. With Shopify, GA4 and Meta Ads connected, "Which campaigns drove the most revenue this month?" becomes calls to meta_ads_get_campaign_insights, ga4_run_report and shopify_get_orders. Joining those numbers is the model's work, so check the answer against a report you trust the first few times.

Common misconceptions about MCP servers

"Once a server is connected, the AI can do anything"

The model can only call tools the server lists. If there is no delete tool, it cannot delete.

The spec also says a host must get the user's consent before it invokes a tool, and notes that the protocol itself cannot enforce this (spec). It is up to the AI app. Claude, for example, asks you to approve a tool call and offers an "Allow always" option, which its help center says to use only for servers and tools you trust (Claude help center).

"One MCP server covers every tool"

A server usually covers one product. Shopify and HubSpot are two servers. A hosted platform can put both behind one sign-in, but each product still has its own tools and its own connected account.

"The vendor's own server is always the best pick"

It is a sensible default, because the vendor knows its API. The vendor also chooses how much of that API to expose. If your job needs an operation the server leaves out, you need a broader server or the API itself. Compare the tool list to the job before you commit. Best MCP servers lists options by use case.

How to get an MCP server

There are three routes.

1. Use the vendor's server. Search the product's docs for "MCP". This fits when you need one product and its tool list covers your job.
2. Build your own. This fits internal systems nobody else supports. The MCP project lists official SDKs for ten languages, with TypeScript, Python, C#, Go, Rust and Ruby at Tier 1 (SDKs), and the MCP Inspector for testing. You own hosting, sign-in and upkeep.
3. Use a hosted platform. This fits teams that need several business tools and do not want to run servers.

The official MCP Registry is a metadata index of public servers, still in preview, that marketplaces and other aggregators pull from. To browse servers yourself, use one of the directories in MCP server directories compared.

The hosted route with PopMCP

You connect a provider account once from the MCP server catalog, such as Shopify or HubSpot. Then you add one URL, https://app.popmcp.com/mcp, to your AI client, sign in and tick the connectors that client may reach. There is no token to paste.

Try a read first. Ask the client to run list_connections so you can see which accounts it reaches, because a write runs as soon as the model calls it. How to connect Claude to your business tools walks through each step, and the Claude setup page has the short version.

Provider credentials stay in PopMCP, encrypted at rest with AES-256-GCM, and are never passed to the AI client.

On catalog size: HubSpot has 1,000+ operations in PopMCP, Klaviyo 308 and Shopify 70. PopMCP loads a curated set for each provider and keeps the rest reachable through a search tool, for example hubspot_search_tools, on every plan.

Where it does not fit: PopMCP covers business tools, so it will not read local files or a Git repo. It is not a workflow builder or an agent framework, and it does not run jobs on a schedule. If a provider you need is missing, request an integration.

Questions to ask before you connect any server

• Who runs it, and when was it last updated?
• Which of its tools write or delete?
• Where does the credential live: in the AI app's config file, on your machine or with the provider?
• How do you cut off access later?
• Does your AI app support its transport? Web clients need a remote server.

MCP security best practices has a fuller checklist.