What is MCP? A plain explanation, no code needed

AI is smart, but stuck inside the chat box

Claude or ChatGPT can answer questions, summarise documents and draft work well, but on its own it cannot see your email, files, shop sales or calendar. Every time you want help, you copy and paste the data in yourself.

MCP fixes that. It is a shared agreement that lets AI apps connect to other software. Once connected, the AI can fetch what it needs and do some tasks for you, within the permissions you give it.

Compared with things you already know

Like apps on a phone

A bare phone can make calls; add a banking app and it can send money. AI apps work the same way: connect a Google Drive server and it can find your files.

How is it different from an API?

An API is each product’s own door, and every door works differently, so developers learn them one by one. MCP is one language the AI uses at every door; the server maker does the translating.

How is it different from old plugins?

Earlier plugins were tied to one app. MCP is an open standard: one server works in Claude, Cursor, VS Code and other apps that support it.

Only three parts

The AI app you use

Called the client, e.g. Claude Desktop, Claude Code, Cursor or VS Code.

The connector

Called the server; each one connects to one service, such as GitHub, Notion or a database.

The buttons AI can press

Called tools, e.g. “search files” or “create invoice”. Each server tells the AI which buttons it has.

On your computer, or in the cloud

Local means the server runs on your computer and your AI app starts it. You usually need Node.js, Python or Docker first. Your data stays close to you.

Remote means the maker runs the server online. You paste its address (the endpoint) into your AI app and sign in. Easier, but what you send passes through the maker’s servers.

Five things before you install

Check who made it

A server from the company behind the service (GitHub’s own GitHub server, for example) is usually more trustworthy. See the “maker matches the repo” check on each page.

Grant only what is needed

If the service lets you choose permissions for an API key, start with read-only.

Keep keys secret

Keys and tokens open your account. Never paste them into a chat or send them to anyone. If one leaks, revoke it and create a new one.

Watch for text that instructs AI

Web pages, emails or files can hide instructions telling AI to send your data out. This is called prompt injection. Have your AI app ask before it sends, deletes or pays.

You can always remove it

When you stop using a server, delete its settings from your AI app and revoke any key you created.

Your first install in 4 steps

Find a server for your task in the registry or the AgentSala list.
Open its page, choose the tab for your AI app and press Copy.
Paste the settings into your AI app, fill in any keys, then restart the app.
Try one of the example requests on that page. If the AI asks to use a tool, read it before you allow it.

Glossary

Every term with a dotted underline on this site is explained here.

MCP (Model Context Protocol): A shared language that lets AI apps connect to other software and data, like a standard socket that fits every brand.
MCP server: A connector between AI and one service, such as GitHub or Notion. It tells the AI what it can do, then does that work.
client: The AI app you use, which connects to servers, e.g. Claude Desktop, Claude Code, Cursor or VS Code.
tool: One action a server lets the AI perform, such as “search files” or “create invoice”. Good AI apps ask you before using one.
registry: The Official MCP Registry is the shared list where makers publish their servers. AgentSala reads its list from there.
local: The server runs on your computer and your AI app starts it. You usually need Node.js, Python or Docker installed.
remote: The maker runs the server online. You just paste its address into your AI app, but your data passes through the maker’s servers.
stdio: How an AI app talks to a server running on the same computer, through an internal channel that never leaves the machine.
Streamable HTTP: How an AI app talks to a cloud server over the internet. The current standard way.
SSE: An older way MCP talked over the internet. Still works, but most makers have moved to Streamable HTTP.
transport: The channel an AI app and a server use to talk: local (stdio) or over the internet (HTTP).
endpoint: The internet address of a cloud server. It looks like a web link, e.g. https://…/mcp
OAuth: A sign-in method where your AI app sends you to the service’s own page to approve access, so you never copy a key around. Safer.
API key: A long code a service gives you so software can use your account. Anyone who has it can act as you, so keep it secret.
token: A kind of password software uses on your behalf. Set it to expire and to allow only what is needed.
environment variable: A setting where you put personal values for a server, such as a key or a folder path, without touching code.
licence: The maker’s terms for who may use, change or resell the software. MIT and Apache-2.0 allow broad use.
GitHub stars: How many people starred the repo on GitHub. It shows interest, not safety or quality. If one repo holds several servers, they share its stars.
repo: Where a project’s code lives on GitHub. You can see who made it, when it last changed and what problems people reported.
basic checks: What AgentSala checks from public information: maker identity, maintenance, licence and endpoint. Not a code scan yet.
uptime: The share of time a cloud server responds. For now we check once; a later step checks on a schedule and reports a percentage.
percentile: A position compared with others. “More stars than 80%” means more stars than 8 out of 10 servers in the same category.
npm: The package library for JavaScript. Servers on npm start with the npx command and need Node.js.
PyPI: The package library for Python. Servers on PyPI usually start with uvx and need uv installed.
Docker: Software that runs programs in a box separate from your computer. Some servers ship as Docker images.
prompt injection: Hidden text in a web page, email or file that tries to make AI do something you did not ask, such as sending your data out.
open source: Software whose code anyone can read, so outsiders can check what it really does.

Where our numbers come from

Every ⓘ next to a number links here.

GitHub stars: Stars on the repo the registry points to, fetched via the GitHub API on Nov 5, 2026. If a repo holds several servers, they share its stars.
Basic checks: Checks passed out of those that apply, run by AgentSala on Nov 5, 2026 from public information. Checks that do not apply are not counted.
Usage level: GitHub stars in words: Widely used (10,000+) · Growing (1,000+) · Early (100+) · Very new (under 100).
User voice: Ratings from people who use the server. Opens with accounts (and account deletion) in a later step.
Version: The latest version number the maker published to the Official MCP Registry.
Published to registry: When the maker published the latest version to the Official MCP Registry, read from registry data on Nov 5, 2026.
Code last changed: The latest push to the repo on GitHub, fetched on Nov 5, 2026.
Endpoint responds: AgentSala sent one MCP initialize request to the endpoint, waiting up to 6 seconds, on Nov 5, 2026. Not a daily uptime figure yet.
Category comparison: Compared only with servers in the same category in this prototype, using data fetched on Nov 5, 2026.
Licence: The licence type GitHub detected from the repo’s files, fetched on Nov 5, 2026.
Category: Assigned automatically by AgentSala from the name, description and repo topics (a few obvious misses fixed by hand). It may still be off.
Number of servers: This prototype shows 100 servers with GitHub repos from the Official MCP Registry (the real registry has tens of thousands), fetched on Nov 5, 2026.
Usable from the cloud: Counted from servers whose maker listed an endpoint in the registry.
Installable locally: Counted from servers whose maker listed an npm, PyPI, Docker or NuGet package in the registry.
Maintained: Counted from servers whose GitHub repo changed within 90 days before the fetch date (Nov 5, 2026).

Keep going

Browse all MCP servers →MCP for online shops