How many MCP servers are there?

There is no single official count, but public directories now list thousands. AllMCPs catalogs more than 3,000 Model Context Protocol servers across over 50 categories, and the number grows every week as developers publish new integrations. The largest categories are developer tools, finance, and knowledge and memory.

What is the difference between an MCP server and an API?

An API is a raw interface you have to read docs for and write integration code against. An MCP server wraps a capability — which might be a REST API, a database, or a local tool — in the Model Context Protocol, so any MCP-compatible AI client can discover its tools and call them in a standard way without custom glue code. One MCP server can also sit in front of several APIs at once.

Are MCP servers free?

The large majority are open source and free to self-host. Some wrap a paid third-party service (a search API, a market-data feed, a cloud provider) and require your own API key or account for the underlying service, but the MCP server itself is almost always free.

Which MCP server should I install first?

Start with something low-risk and immediately useful, like the official filesystem or fetch server, or a read-only database server pointed at a non-production database. Each listing on AllMCPs shows the exact install command, so you can copy a config and have your first server running in a couple of minutes.

What Can MCP Servers Do? A Field Guide to the Model Context Protocol Ecosystem

TL;DR: Model Context Protocol servers are the plug-ins that give AI agents real capabilities — querying databases, searching the web, driving a browser, reading your files, calling APIs. There are already more than 3,000 of them across 50+ categories. This guide explains what they do, how the ecosystem breaks down, how local and remote servers differ, and how to pick the right one instead of drowning in options.

If you have spent any time with Claude, Cursor, or another AI coding tool lately, you have probably run into the phrase "MCP server." The Model Context Protocol has quickly become the standard way to give an AI agent new abilities, and the number of available servers has exploded. But "MCP server" is a broad label — a server that reverse-engineers binaries and a server that books restaurant tables are both MCP servers. This guide is a map of the whole landscape: what these servers actually do, how many there are, and how to choose one.

What an MCP server actually is

An MCP server is a small program that exposes a set of capabilities — tools the model can call, resources it can read, and prompts it can reuse — over the Model Context Protocol. Your AI client (Claude Desktop, Claude Code, Cursor, Windsurf, VS Code, and others) launches or connects to the server, discovers what it offers, and can then use those capabilities during a conversation.

The key idea is standardization. Before MCP, every integration between an AI app and an external tool was bespoke. MCP replaced that with one protocol, so a server written once works in every compatible client. If you want the deeper conceptual explanation, we cover it in What is the Model Context Protocol? — this article focuses on the ecosystem that has grown on top of it.

How big is the MCP ecosystem?

Large, and growing fast. The AllMCPs directory alone catalogs more than 3,000 active MCP servers across 50+ categories, and new ones arrive every week. That growth is not evenly spread — a few categories dominate, which tells you a lot about what people are actually building for.

Here is roughly how the catalog breaks down at the time of writing:

Category	What these servers do	Explore
Developer Tools	Run code, manage repos, query build systems	/categories/developer-tools
Finance & Fintech	Market data, payments, accounting, on-chain activity	/categories/finance-and-fintech
Knowledge & Memory	Persistent memory, note stores, searchable knowledge bases	/categories/knowledge-and-memory
Security	Scanning, auditing, secrets management, threat analysis	/categories/security
Search & Data Extraction	Web search, scraping, structured extraction	/categories/search-and-data-extraction
Databases	Query and inspect SQL and NoSQL databases	/categories/databases
Communication	Slack, Discord, email, and other messaging	/categories/communication
Browser Automation	Drive a real browser to navigate, click, and extract	/categories/browser-automation

Developer tools is the single largest category, which makes sense — the first people to adopt MCP were engineers wiring it into coding agents. But finance, knowledge and memory, and security are all substantial, a sign that MCP has spread well beyond writing code.

What MCP servers can do, by capability

Rather than list 3,000 servers, it helps to think in terms of what they let an agent do. A handful of capability groups cover most of the ecosystem.

Reach your data

Some of the most-installed servers connect an agent to data you already have. Database servers let an agent run queries and inspect schemas — PostgreSQL, SQLite, and MySQL are the most common. Knowledge and memory servers give an agent persistent recall and searchable notes across sessions. And filesystem servers give scoped read and write access to local files.

Reach the web

Search and data-extraction servers let an agent search the web and pull clean, structured content out of pages. Browser automation servers go further — they drive a real browser with tools like Playwright so an agent can navigate, fill forms, and interact with sites that have no API at all.

Reach your tools and platforms

This is where the long tail lives. There are MCP servers for GitHub, Docker, cloud platforms, YouTube transcripts, Notion, Slack, Jira, Stripe, and thousands of other services. Many of them are generated from an existing API — OpenAPI-to-MCP servers can turn any documented REST API into MCP tools automatically, which is a big reason the catalog grows so quickly.

Do specialized work

Beyond the mainstream categories, the ecosystem gets delightfully specific: servers for reverse-engineering binaries, for astronomy calculations, for game engines, for legal research, for multimedia processing. If a task has an API or a command-line tool, someone has probably wrapped it in MCP.

Local vs remote: how MCP servers run

MCP servers come in two deployment shapes, and knowing the difference matters for both setup and security.

Local (stdio) servers run as a process on your own machine, launched by your client on demand (usually via npx or uvx). They are the default for developer tooling because they can touch your local files, shell, and git checkout. Most of the catalog is stdio.
Remote (HTTP/SSE) servers run somewhere else and your client connects over the network. These are how hosted and multi-user services expose MCP, and they introduce real authentication and authorization concerns — which we cover in depth in Securing remote MCP servers.

A single listing often supports both, and each server's page on AllMCPs shows the transport and a ready-to-paste config for either.

How to choose the right MCP server

With thousands of options, the hard part is not finding an MCP server — it is picking a good one. A few criteria do most of the work:

Official vs community. An "official" server is published or claimed by the project it integrates with, which usually means better maintenance and trust. Community servers are often excellent too, but check activity before you rely on one.
Maintenance signals. Look at GitHub stars, recent commits, and whether the repo is archived. A popular, recently updated server is a safer bet than a one-off from two years ago.
Scope of access. Prefer the least-privilege option. A read-only database server is safer than one with write access; a browser server scoped to headless mode is safer than one with your logged-in profile.
Transport fit. Need local file access? You want a stdio server. Building a hosted product? You want a remote server with real auth.
Popularity for your exact use case. Our best-of guides rank servers by real usage for specific needs — databases, GitHub, browser automation, and more — so you can start with what the community actually relies on.

If you are weighing two specific options head to head, our guide to comparing MCP servers walks through doing it well.

Getting started

Once you have picked a server, installing it is quick and nearly identical across clients — you add a small JSON block and restart. Our step-by-step install guide covers Claude Desktop, Claude Code, Cursor, Windsurf, and VS Code, and if a server connects but shows no tools, the troubleshooting guide will get you unstuck.

From there, the best way to understand the ecosystem is to explore it. Browse the full directory, filter by category, or start from a best-of ranking for the job you have in mind. And if you maintain an MCP server yourself, you can add it to the directory for free so others can find it.

The Model Context Protocol turned "AI that can chat" into "AI that can act," and the server ecosystem is where that capability actually lives. Three thousand servers is a lot — but with a map of what they do and a few criteria for choosing, finding the right one is the easy part.