Understanding & Using MCPs

An MCP is an extra capability for your AI. It lets it not only answer but also look things up for you and act on your behalf – in your notes, your project board or the web. Setting one up takes a few clicks, and you don't need to know anything about how it works inside. Buttons and tabs are named exactly as the app shows them. Where the app's wording is German, the English meaning follows in brackets.

On This Page

What Is an MCP?
What MCPs Are Good For
What an Extension Needs
Adding an MCP
Signing In to a Service
Using an MCP in Chat
Turning On and Off
When Something Doesn't Work
Security in a Few Words

What Is an MCP?

By default your AI can read, write and answer you. What it doesn't know is your world: your notes, your tasks, yesterday's website. That is exactly what an MCP changes (short for "Model Context Protocol"). You connect Apis AI to a service once – afterwards the AI may use its information and get things done for you there.

Think of your phone: the device alone can make calls. Apps turn it into a camera, a navigation system or a note collection. MCPs are such apps – just for your AI.

In Short Without an MCP the AI answers from its own knowledge. With an MCP it can look things up for you and enter things on your behalf. You don't have to program anything: MCPs are managed in the app, and in chat the AI triggers them itself when you describe your goal.

What MCPs Are Good For

In the MCP-Bibliothek (MCP library) you'll find ready-made extensions for the most common tasks. A few examples of what you can achieve with them:

What you want to achieve	Matching extension
Summarize a website without opening it	fetch, Firecrawl
Research on the web and collect sources	Brave Search, Exa, Tavily
Store and find notes and knowledge	Obsidian, Notion, memory
Maintain tasks in your project board	Linear, GitLab
Keep an eye on errors and issues	Sentry
Understand what a public project does	DeepWiki, context7
Click in the browser and take screenshots for you	Playwright, Chrome DevTools
Prepare posts for social networks	Buffer
Query figures from a database	postgres
Include specific folders on your machine	filesystem

These are only examples from the curated list. In the Registry (online) tab you'll find many more extensions – including for niches that are not listed here.

What an Extension Needs

Most extensions work as soon as you add them. A few want something first: a small program on your computer or an access key from the service. The description in the library tells you which applies – here is the overview:

Extension	What it needs
context7, DeepWiki	Nothing – ready right after adding
Hugging Face	Nothing – an access key only if you want to use private models
fetch, memory, sequential-thinking	Node.js on your computer
filesystem	Node.js and the folder you want to allow
Playwright, Chrome DevTools	Node.js on your computer plus a browser (Chrome or Chromium)
Brave Search, Exa, Tavily, Firecrawl, Notion	An access key from the respective service
Buffer	An access key from Buffer
Obsidian	Node.js, the Obsidian plugin Local REST API and an access key
Linear, Sentry, GitLab	Signing in with your account
postgres	The access details of your database (user, password, host, database) and the helper program uv

Helper programs like Node.js or uv are free – you install them once and several of these extensions then work. If you would rather not install anything, simply pick an extension that manages without. Where an access key is needed, you get it from the service itself, usually in your account there, and enter it in the card in the Environment variables field – one line per entry, in the form Key=Value (with Buffer and Hugging Face it goes into the Headers field instead). Where "signing in" is written, the Anmelden button on the card is all you need.

Tip If you would rather not set anything up at all, start with an extension that connects directly to a service on the internet – those are ready to go the fastest.

Adding an MCP

Open settings Click Settings in the sidebar and switch to the Erweiterungen tab (Extensions).
MCP Server card In the Erweiterungen tab you'll find the MCP Server card. Click MCP-Bibliothek (MCP library) there.
Choose an extension Under Kuratiert (Curated) you'll find hand-picked extensions. In the Registry (online) tab you can also search. The plus on an entry adopts it – the symbol then turns into a check mark.
Check the card The extension appears as its own card. There you also decide whether it runs on your computer (Local (stdio)) or connects to a service on the internet (Remote (HTTP)). If a placeholder is still there – such as an example path or a pattern for username and password – replace it with your own details.
Save settings Click Save Settings. Only then does the app connect – the status appears right below the name of the card.

Tip for Getting Started Start with an extension that only reads something, for example fetching a website. Those are the easiest to set up and you see the effect in chat straight away.

Important Some extensions contain placeholders such as an example folder or a pattern for username and password. As long as you don't replace these with your own values, the connection stays off.

Not in the Library? If a service is missing, use Add an MCP Server on the same card. That creates an empty card which you fill in yourself – it needs details such as an address or a command. The individual steps are in Setting Up MCP Servers.

Signing In to a Service

Some services – such as your project board or your error tracking – want you to sign in first. After adding, the card then shows Anmeldung erforderlich (sign-in required) and the button Anmelden (Sign in).

Click Anmelden (Sign in) Your browser opens and asks you to sign in to that service.
Log in and confirm Sign in with the account that may use this service and approve the access.
Done The status changes to Linked (… Tools) – the extension is ready to use.

Your sign-in stays stored on your machine. If you no longer want it, remove the card – that also removes the connection.

Using an MCP in Chat

You don't "call" an extension. You simply describe what you want – and if a matching extension is active, the AI uses it on its own. For example:

"Summarize this page for me: …" (extension for reading websites)
"Find me three current articles on the topic …" (extension for web search)
"Add a task for next week to the project board." (extension for your board)
"Look in my notes for what I've recorded about …" (extension for notes)

How to check whether something was really looked up: Under the response it says Tool calls (…). Clicking it shows you which extension was used and what came out of it. Very generic or made-up answers are often a sign that the matching extension is not active.

Tip The more concretely you describe your goal, the more reliably the right extension kicks in. Instead of "Help me with the project", prefer "Create a task 'Prepare the move' and set the deadline to next Monday".

Turning On and Off

You stay in control at all times. On every server card a badge below the name shows you the status:

Status	What it means
Linked (… Tools)	Ready – the extension can be used in chat. The number tells you how many individual capabilities it brings along.
Verbinde… (Connecting…)	The app is currently establishing the connection.
Anmeldung erforderlich (sign-in required)	You still need to click Anmelden.
Connection Failed	The service was not reachable. Next to it you usually see a short error message – check the details and try again.
Disabled	Switched off with the switch on the card.
Not connected yet	The server has only just been added and has not connected yet.

With the switch on the card you switch a server off without deleting it – the AI can then no longer use it. The sidebar always shows how many servers are active. In workflows you can additionally decide per step which servers are allowed there (see Workflows). And in the tab Extensions you decide per model which servers are sent along at all – the card Tool Selection per Model keeps models with a small context window usable (see Settings).

When Something Doesn't Work

It stays at "Anmeldung erforderlich" – click Anmelden and complete the sign-in in the browser window. Only then does the extension connect.
"Connection Failed" – usually a placeholder has not been replaced, a path points to a folder that doesn't exist, or credentials have expired. Correct the card and save the settings again.
"Command not found" / "Kommando … wurde nicht gefunden" – the extension needs a program that is missing on your machine: extensions that run with npx need Node.js, ones with uvx need uv. Install the program and then use Erneut verbinden (Retry) in the notice at the top of the chat. If the program is installed and still not found, put its full path into the Kommando field of the server card (see Setting Up MCP Servers).
A notice appears at the top of the chat – Apis AI lists there which MCP servers could not be started and why. Erneut verbinden (Retry) tries exactly those servers again, the × hides the notice until the next attempt.
Nothing happens after a change – changes only take effect once you click Save Settings. If a server then still shows Connection Failed, use the small refresh icon next to the status to try that one connection again.
The AI doesn't use an extension – check three things: Is it active and connected? Is your request concrete enough? Does the workflow step allow this extension?
The chat says "Tool calls disabled" – your model's context window is too small for the extensions you have switched on. This only affects the first message of a chat and the app tells you right there. Either use a model with a larger context window or switch off the servers you don't need for that chat.
The extension needs extra software or an access key – not every extension runs right away. The overview of what each one needs is in What an Extension Needs.

Security in a Few Words

An extension can only do what its purpose is. So only connect services you know.
Your credentials stay on your machine – they are not written into your chats.
What you switch off, the AI cannot use. Removing it also removes the connection.
For important tasks it's worth a look at the tool calls under the response: there it says what the AI actually did.

Rule of Thumb Let it read first, then let it write: start with an extension for looking things up and only then add one that changes or creates something for you.