MCP server MIT
Apify MCP Server
by Apify
Apify MCP server: lets AI agents run Apify Actors to scrape and extract web data, with docs and storage tools.
- Does
- 5 things
- Install
- 5 steps
- Works with
- 3 hosts
Guide
What is an MCP server? A plain explanation for 2026
MCP stands for Model Context Protocol. An MCP server is a small program that exposes a service to an AI app. The AI app is the client, also called the host; examples in our profiles are Claude Code, Cursor, Windsurf and Codex. The client connects to the server and the server tells it what it can do, so one server such as the GitHub MCP Server can work in many clients.
A server offers tools (actions the model can call, such as creating an issue or running a query), and the reference everything server also exercises prompts and resources (data the client can read). Servers connect in two ways. Local servers run on your machine as a command, for example npx for the filesystem server or uvx for git and fetch. Remote servers sit behind a URL, such as Context7, Sentry and Vercel.
Security matters because a server acts with the access you give it. Profiles cite concrete advice: grant minimum token scopes (GitHub), use a Restricted API Key (Stripe), create read-only database users (Redis) and review what the model asks to do (Neon). The fetch server's README warns it can reach local and internal IP addresses, and the filesystem server's write_file overwrites existing files. We profiled 49 MCP servers; 47 are official.
- Official 47
How it works
- The client reads a config (for example mcp.json in Cursor or mcp_config.json in Windsurf) listing which servers to start or connect to.
- Local (stdio) servers are launched as a command such as npx or uvx; remote servers are reached by URL, often with OAuth or an API key in a header.
- The server advertises its tools; the model decides when to call one, and the client runs it and returns the result.
- In Claude Code you can add one with a command like claude mcp add --transport http, as the Zapier profile shows; Codex uses codex mcp add.
- Limit exposure: GitHub supports --toolsets to expose fewer tools, and AWS API offers READ_OPERATIONS_ONLY=true.
MCP vs MCP server vs MCP client
| Term | What it is | Example |
|---|---|---|
| MCP | The protocol: JSON-RPC 2.0 messages plus two transports | The specification at modelcontextprotocol.io |
| MCP host | The AI app you work in | Claude Code, Claude Desktop, VS Code, Cursor |
| MCP client | One connection the host opens per server | VS Code opens a separate client for Sentry and for the filesystem server |
| MCP server | The program that provides tools, data and prompts | GitHub, Sentry, filesystem |
Is an MCP server a real server? Local vs remote
Not always a networked one. A local server is a process the AI app launches on your machine and talks to over standard input/output (stdio); it usually serves one client and runs with your user permissions. A remote server runs on the vendor's platform behind a URL over Streamable HTTP, serves many clients and authenticates with bearer tokens or API keys; the protocol recommends OAuth.
MCP server vs API
An API is one service's own interface, and every AI app has to integrate it separately. An MCP server wraps that capability in a standard format the model can discover at runtime (tools/list) and call (tools/call), so one server works in every MCP host. Many MCP servers call the vendor's API underneath.
What a server can offer
- Tools: actions the model can call, such as running a query or creating an issue
- Resources: data the app can read, such as a file or a database schema
- Prompts: reusable templates, such as few-shot examples
- Elicitation: the server can ask you for missing input. Roots, Sampling and Logging are deprecated as of protocol version 2026-07-28 (SEP-2577): they still work during the deprecation window, but new servers should not rely on them
Examples to start with
Atlassian Rovo MCP Server
by Atlassian
Atlassian's official remote MCP server connecting Jira, Confluence and more to AI assistants with OAuth.
- Does
- 5 things
- Install
- 4 steps
- Works with
- Any MCP client, Cursor
MCP server Apache-2.0
AWS API MCP Server
by AWS Labs (Amazon Web Services)
MCP server that lets AI assistants run AWS CLI commands across all AWS services with validation and security controls.
- Does
- 5 things
- Install
- 5 steps
- Works with
- 3 hosts
MCP server Apache-2.0
AWS Documentation MCP Server
by AWS Labs (Amazon Web Services)
MCP server that fetches, searches and recommends current AWS documentation pages as markdown for AI assistants.
- Does
- 5 things
- Install
- 4 steps
- Works with
- Any MCP client, Cursor
MCP server Apache-2.0
AWS MCP Servers (awslabs/mcp)
by AWS Labs (Amazon Web Services)
Open-source suite of MCP servers from AWS Labs for AWS docs, APIs, IaC, pricing, data and more, run locally over stdio.
- Does
- 5 things
- Install
- 5 steps
- Works with
- 4 hosts
MCP server MIT
Azure MCP Server
by Microsoft
Microsoft's MCP server that brings Azure context and operations into AI agents, installable via npm, .NET, PyPI or Docker.
- Does
- 5 things
- Install
- 5 steps
- Works with
- Any MCP client, Claude Code
MCP server MIT
Brave Search MCP Server
by Brave
Brave's official MCP server for web, local, image, video and news search plus AI summarization.
- Does
- 5 things
- Install
- 4 steps
- Works with
- Any MCP client
MCP server Apache-2.0
Chrome DevTools MCP
by Chrome DevTools (Google)
Lets your coding agent control and inspect a live Chrome browser: performance traces, network, console, screenshots, automation.
- Does
- 5 things
- Install
- 4 steps
- Works with
- 3 hosts
Frequently asked questions
What does MCP mean?
MCP means Model Context Protocol, the standard that lets AI apps connect to external tools and data through servers.
What is the difference between stdio and remote MCP servers?
A stdio server runs locally as a command like npx or uvx. A remote server runs elsewhere and you connect to its URL, usually with OAuth or a token.
Do I need to code to use an MCP server?
Not usually. You add a config entry or run an add command and authenticate. To build your own, Anthropic's mcp-builder-skill and mcp-server-dev plugin guide you.
Are MCP servers safe?
They are as safe as the access you grant. Use minimum permissions, read-only credentials where possible, and read each server's README security notes.
What is the difference between an API and an MCP server?
An API is one service's own interface that each AI app must integrate separately. An MCP server wraps that capability in a standard format the model discovers at runtime and calls, so one server works in every MCP client; many MCP servers call the vendor API underneath.
Is an MCP server a real server?
Not always a networked one. A local MCP server is a process the AI app launches on your machine over stdio; a remote server runs on the vendor's platform behind a URL, often with OAuth.
Does ChatGPT use MCP?
Yes. The Model Context Protocol site lists Claude, ChatGPT, VS Code and Cursor among the clients that support MCP servers.