Connectors: what access you give

This lesson is still being tested: it improves with learners' feedback. If something is unclear, it is the lesson that needs fixing, not you.

By the end of this lesson, you will be able to

say what a connector is, and what it opens up to Claude;
ask the four questions before adding one;
add, use and then remove a read-only connector in Claude Code;
apply the rule: read-only first.

Why it matters to you

A connector gives Claude access to another service: your email, your calendar, your online documents, a work tool. It is very useful, and it is the moment when you open a door. Well chosen, it saves you hours; badly chosen, it lets the AI read or edit more than necessary. This lesson teaches you to open the right door, the smallest possible one, and to close it again.

Before you start


You need	Claude Code in the terminal (module 4, lesson 2), in Manual mode, and your mon-espace-ia folder
Cost	The agent subscription: Claude Pro at minimum, €18 per month excl. VAT, price shown in France on 24/10/2026. Some connected services have their own cost
Duration	40 minutes
Reminder	The exercise connector only reads the course’s public lessons. The words of your searches are sent to it: nothing personal in them

Understand it in 3 minutes

A connector is a bridge. Technically, it is called an MCP server. According to Claude’s help, a connector lets Claude access your apps, find your data in them and act in them, with your rights in that service. It cannot go beyond what your account allows, but it can go all the way there.

One exception, according to the same help: a custom connector that uses a shared key accesses everything that key allows. And a server launched by a command is a program on your computer (see the third question below).

Four questions before adding one:

Question	Why
What do I really need?	Reading your appointments is not creating them. The need tells you the smallest door that is enough
What can it do?	Read only, or also write, send, delete? Look at the list of its tools
Who made it, and who do I trust?	The Claude Code documentation says it: check that you trust each server before connecting it; a server that fetches outside content can expose you to hidden instructions. And a server that starts with a command, and not with an internet address, runs on your computer, with direct access to your machine: never paste such a command without understanding it
How do I remove it?	You must know how to close the door before you open it

Read-only first. Start with a connector that only reads, or with only the read tools of a connector. You will add writing later, if a real need appears.

Where connectors are set up:

in claude.ai and the app: in Customize, in the Connectors section, or with the + button of a conversation;
in Claude Code: with the claude mcp command, in the terminal. According to the documentation, a connector added with the --scope project option is written to a .mcp.json file in your folder, and only works in that folder.

The two meet. If you are signed in to Claude Code with your claude.ai account, the connectors you added on claude.ai (Gmail, Calendar, Drive…) are automatically available in Claude Code too (Claude Code documentation, 24/10/2026), and Anthropic may add one itself, without you doing anything (Claude Docs, for writing documents). To see everything that is connected, type /mcp in a terminal session: connectors that come from claude.ai are marked there, and you can turn them off there for this folder only. To keep them out everywhere, disconnect them on claude.ai: Customize, then Connectors, then Disconnect.

Do it together

The exercise connector is the course server: it only reads the public lessons. It stores nothing personal: step 1 says what it receives, and what the site’s host keeps.

1Step 1: Ask the four questions

Need: search the course lessons from Claude Code.
What it can do: six tools, all read-only (catalog, read a lesson, an exercise, a resource, the next lesson, a search).
Trust: it is published by the course website; it receives the identifier of the lessons or resources requested, the words of your searches and, for the next lesson, the list of lessons already done. As with any website, its host keeps technical logs of visits, including the IP address.
Removal: one command, in step 4.

2Step 2: Add it, for this folder only

In the terminal, go to your mon-espace-ia folder (module 4, lesson 2), then type:

claude mcp add --transport http --scope project cours https://neurylio.com/api/mcp

The command replies that it added the cours server to the project configuration. A .mcp.json file has appeared in your folder. Check:

claude mcp get cours

As long as you have not approved it, it is shown as “Pending approval”.

3Step 3: Approve it, then use it

Start Claude Code in Manual mode (if your version is too old and refuses “manual”, update Claude Code, or use claude --permission-mode default, which starts in the same mode):

claude --permission-mode manual

According to the documentation, Claude Code asks you to approve a server declared in .mcp.json before using it. Read the request, then approve. Next, write:

With the cours connector, search the English version of the course for the lessons that talk about memory. Give only their titles.

A permission request should appear for the connector’s search tool (this request could not be shown on screen during testing): read it, then allow. You get a list of lessons, including “Automatic memory: view, correct, turn off”.

4Step 4: Remove it, and check

Quit Claude Code (/exit), then:

claude mcp remove cours -s project

The command confirms the removal. Check that claude mcp get cours no longer finds it. You now know how to add a connector, limit it to one folder, and remove it.

Your turn

A sheet for a real connector that you would like (calendar, online documents, email, work tool), without installing it:

answer the four questions;
look in its presentation page for what it lets you read, and what it lets you do;
decide: add it as read-only, add it later, or do not add it;
write your decision and its reason in your journal.

Check your work

It worked if:

you added the course connector for your folder only, used one of its tools after a permission request, then removed it;
claude mcp get cours no longer finds it;
your sheet for a real connector answers the four questions, with a decision and its reason.

Common mistakes:

adding a connector “just to see”, with all its write tools;
allowing once and for all a tool that sends or deletes;
forgetting that a connector acts with your rights: if it can read your email, so can the AI;
leaving unused connectors: they keep an access open.

Keep it and pick it up later

Write down in your journal every connector you add: when, why, with which rights, and how to remove it.
Once a month, during the upkeep of your space (module 5, lesson 7), remove the ones you no longer use.

Going further

optional

The official documentation, in English: code.claude.com/docs/en/mcp (installation, scope of connectors, approval, removal).
In claude.ai and the app, Claude’s help describes the connectors directory and how to add them: support.claude.com, “Use connectors to extend Claude’s capabilities”.

New words

4 words explained

Word	What it means
Connector	A bridge between Claude and another service, which gives it access to your data and to some actions
MCP server	The technical name of a connector in Claude Code
Read-only	An access that lets you read, without writing, sending or deleting
`.mcp.json`	The file in your folder that lists the connectors of this project

Adding, using with permission, and removing the course connector were actually run with Claude Code 2.1.310 in a test folder (test report, T39). The interactive approval of the server was not run (the test goes through non-interactive mode, where it does not appear).