Safe connections and emergency stop
This lesson is still being tested: it improves with learners' feedback. If something is unclear, it is the lesson that needs fixing, not you.
Follow this lesson with Claude Code
With the course companion installed, open Claude Code in your working folder and type:
/formation-ia:suivre m08-l06 enIt reads the up-to-date lesson on this site and guides you step by step, without changing anything without your consent. Install the companion. Claude Code is not included in the free plan: an eligible Claude subscription, or a Console account billed by usage for the terminal.
By the end of this lesson, you will be able to
- choose a safe connection: official method, minimal rights;
- reduce an automation’s access to what it needs;
- prepare an emergency stop plan;
- try it out in calm conditions, so that it works the day you need it.
Why it matters to you
Sooner or later, an automation will do something unexpected: a run at a strange time, an action you hadn’t planned, a forgotten access. On that day, the difference comes down to one minute: knowing what to cut, and where. This lesson prepares for that moment, in calm conditions.
Before you start
| You need | The “My extensions” list in your journal (module 7), and the stop plan (resource) |
| Cost | The agent subscription: Claude Pro at minimum, €18 per month excl. VAT, price shown in France on 24/10/2026 |
| Duration | 35 minutes |
| Reminder | Never paste a password into a conversation to “connect” a service |
Understand it in 3 minutes
A safe connection is:
| Rule | In practice |
|---|---|
| The official method | The connector from Claude’s directory, or the publisher’s procedure; never a password pasted into a conversation |
| Minimal rights | Read if reading is enough. According to Claude’s help, a connector acts with your rights in the service: every right that is open can be used |
| Only the useful connectors | In a cloud task, the documentation says that all your connectors are included by default, writes included: remove the ones it doesn’t need |
| Approval of actions that commit | In your conversations on claude.ai and in the app, the connector’s settings (the Tool permissions section) let you set the tools that send, share or delete to Needs approval or Blocked: never Always allow. Warning: Claude Code’s documentation does not say that these settings apply to its sessions (routines, scheduled tasks), and a cloud task normally starts in auto mode: a control model allows or blocks, in your place, most of the actions that would have waited for your approval. For these tasks, disconnect the connectors they don’t need rather than counting on an approval request |
Where to cut, in case of a problem (according to the help pages and the documentation, 24/10/2026):
| What is acting | Where to cut it |
|---|---|
| A local scheduled task | App, Code tab, Scheduled: status Paused, or delete the task |
| A cloud task | claude.ai/code/routines: open the task, then pause it with the on/off switch at the top of its page |
| A claude.ai or app connector (Gmail, Calendar, Drive…) | Customize, then Connectors: choose the connector, then Disconnect. Claude is then no longer connected to the service, anywhere. The + button in a conversation only cuts it in that conversation: it stays connected |
A server added in Claude Code (claude mcp add) | /mcp to disable it in this project, claude mcp remove to remove it. For a connector that comes from claude.ai: in the terminal, /mcp only cuts it for the current project; in the desktop app, no Claude Code setting reaches it. To stop it everywhere, disconnect it on claude.ai, line above |
| Hooks | "disableAllHooks": true in your personal settings file ~/.claude/settings.json. Also check that the project’s .claude/settings.json file doesn’t say otherwise: it may take precedence over yours. Prepare this move in calm conditions |
| Plugins (and what they start: hooks, servers, background tasks) | claude plugin disable --all in the terminal, then claude plugin list to check that none is still enabled; then close your running sessions: disabling only applies to the next sessions |
| Claude’s access to your Google account | myaccount.google.com/connections (the linked apps page of your Google account): look for “Claude”, open it, then choose “Delete all connections” (label as written in Claude’s help) |
Cutting does not erase the past. According to Claude’s help, the data retrieved by the Google connectors stays with the conversation where it was retrieved: delete the conversation if needed (it disappears from your history right away, and Anthropic says it erases it from its servers within 30 days).
The runs of a scheduled task are sessions: deleting the task archives them without erasing them, so delete them too.
Do it together
1Step 1: Take an inventory of the access
Take your “My extensions” list (module 7) and add your module 8 connections to it: connectors, scheduled tasks. For each one: what it can read, what it can do, and whether it really needs it.
2Step 2: Reduce
For each access:
- useless: remove it (module 7, lesson 7);
- too broad: for a Claude connector used in your conversations, open it (Customize, then Connectors) and, under Tool permissions, set to Blocked the tools that send, write, share or delete that your task doesn’t need; for a routine, disconnect the connectors it doesn’t need instead; for another service, replace it with a narrower access, if the service allows it;
- cloud task: remove the connectors it doesn’t use.
3Step 3: Fill in the stop plan
Open the Emergency stop plan resource and fill in part 1 with your automations: for each one, the exact place to cut it.
4Step 4: The stop drill, in calm conditions
Time yourself: following part 1 of the plan, pause your point-semaine-auto routine and disable a connector. Under 2 minutes? Your plan works. Then turn everything back on.
Your turn
The module 8 deliverable: your supervised automation (lessons 4 and 5), with:
- its complete task sheet;
- the list of its access, reduced to the minimum;
- your stop plan filled in, tried once in calm conditions, in under 2 minutes.
Resources to download
- Emergency stop plan: cutting in 2 minutes, removing access, assessing and repairing, restarting carefully.
Check your work
It worked if:
- each access of your automations has a reason to exist, and the others are removed;
- your stop plan says, for each automation, where to cut it;
- the stop drill in calm conditions took under 2 minutes;
- you turned everything back on after the drill, and checked that the routine runs.
Common mistakes:
- a stop plan that was never tried;
- keeping connectors “just in case”;
- looking for the cause before cutting;
- forgetting the data already retrieved in conversations.
Keep it and pick it up later
- Keep the stop plan where you can find it in 10 seconds, even if your computer has a problem: printed, or on your phone.
- Update it with every new automation.
Going further
optional
- In your Google account, the linked apps page shows you the apps from other developers that you have given access to your account, not only Claude: it is a good time to sort them out.
New words
3 words explained
| Word | What it means |
|---|---|
| Minimal rights | Giving a tool only the access it needs, nothing more |
| Stop plan | The list, prepared in advance, of what to cut and where, in case of a problem |
| Drill in calm conditions | Trying the stop plan when everything is fine, so that it works the day you need it |
Check your understanding
3 questions, no grade. An explanation appears after each answer.
What's next?
Module 8 is finished. A possible next step: module 9 (“Choosing several AIs and making them work together”), which starts with the lesson “Roles: thinking, researching, producing, reviewing”.
Module 8:
✏️ Do the exercisesMy space: your progress, your exercises and your last lesson.
To be checked: the stop drill in calm conditions has not been tried in the app.
Sources and checks
8 sources
Date, version, sources Written on 24/10/2026. Checks on 24/10/2026:
- connectors (user’s rights, access through the + button and the settings): support.claude.com, “Use connectors to extend Claude’s capabilities”;
- approval by default before sending, sharing and deleting, data kept with the conversation: support.claude.com, “Use Google Workspace connectors”;
- deletion of a conversation (removed from the history right away, erased from the servers within 30 days): privacy.claude.com, “How long do you store my data?”;
- tool permissions of a connector (Always allow, Needs approval, Blocked): claude.com/docs/connectors/getting-started, read on 25/10/2026;
- Claude Code connectors (/mcp,
claude mcp remove): code.claude.com/docs/en/mcp; - cloud tasks (connectors included by default, claude.ai/code/routines page): code.claude.com/docs/en/routines;
- local scheduled tasks (Paused, Delete): code.claude.com/docs/en/desktop-scheduled-tasks;
disableAllHooks: code.claude.com/docs/en/hooks-guide;- apps associated with the Google account: Google’s help (page in French);
- price of Claude Pro: claude.com/pricing.
Sources checked:
- support.claude.com/en/articles/11176164-use-connectors-to-extend-claude-s-capabilities
- support.claude.com/en/articles/10166901-use-google-workspace-connectors
- code.claude.com/docs/en/mcp.md
- code.claude.com/docs/en/routines.md
- code.claude.com/docs/en/desktop-scheduled-tasks.md
- code.claude.com/docs/en/hooks-guide.md
- support.google.com/accounts/answer/13533235?hl=fr
- claude.com/pricing