Skip to content

Chapter 16

Inside an agent: tools, permissions and connections


The chapter in a minute

  • A harness is the software around a model that turns it into a working agent. It holds the instructions, connects the tools, applies the permissions and keeps a log of each step.
  • Every connection is a way in. Ask what it may see and what it may do, give it only what today's job needs, and switch off what you no longer use.
  • Many assistants have a setting for how often they ask first. For anything that sends, buys, deletes or publishes, choose a setting that asks first.
  • MCP is one standard way to connect tools, and every connection made with it is still a permission.
  • Disconnecting stops future access, but it usually doesn't delete what was already shared.

Try it now

Permissions check

0 of 7 steps done

Step 1 of 7

Open the settings of one assistant or account and find its list of connected apps. Assistants and accounts usually keep one. If you can't find it, the companion page has help.

Progress is saved in this browser only.

Try this

Try this, from the book

Find the approval setting in one assistant you use and write down what it is set to now. If you decide to change it, write the old setting down first, so you can put it back.

Tools and settings, as of a date

The book points to the ‘Connections and permissions’ section of the companion page. Here it is in full, with the date each detail was checked. Open it on the companion page

Chapter 16 points here.

Checked: 2 November 2026. Everything in this section was checked again on that date unless a different date is given.

The standard behind many connections

Many assistants connect to your apps through a shared standard called the Model Context Protocol, which you may see named in a settings screen. You don't need to know it to use the book.

Approval settings

ProviderThe levels and labelsWherePage
OpenAI (ChatGPT)Four levels: "Always ask" (asks before reading or changing anything), "Allow read actions", "Allow low-risk actions", and "Allow all actions", which the page says "carries elevated risk". The account-wide setting doesn't offer the last one; it can appear for a single app.Settings, then Plugins (or Apps), then PermissionsManaging app permissions in ChatGPT (opens in a new tab), checked 2 November 2026
OpenAI (ChatGPT Work cloud browser)Website access: "Always ask", "Auto approve" (asks only if a site seems unsafe) or "Always allow", which OpenAI says is not recommended. This is separate from approving important actions.Settings, then Cloud browserUsing cloud browser in ChatGPT (opens in a new tab), checked 2 November 2026
Anthropic (Claude)Per tool: "Always allow", "Needs approval" or "Blocked". The page describes this for Team and Enterprise owners setting rules for everyone. As of 2 November 2026 it still doesn't say whether personal users get the same control.Customize, then ConnectorsUse connectors to extend Claude's capabilities (opens in a new tab), checked 2 November 2026
Anthropic (Claude in Chrome)"Manually approve", "Automatically approve" or "Skip all approvals".The side panel's mode settingClaude in Chrome permissions guide (opens in a new tab), checked 2 November 2026
GoogleA separate list of agent permissions in your Google Account, which you can filter with "Agent access"Your Google Account's linked apps pageManage links between your Google Account and apps from other developers (opens in a new tab), checked 2 November 2026

OpenAI's page also says actions get extra review when they can affect another service, expose sensitive information or be difficult to undo, such as sending messages, editing or deleting files, changing sharing or security settings, purchases and sharing sensitive personal details.

To set your default in ChatGPT (source (opens in a new tab)):

  1. Open your profile menu and select Settings.
  2. Select Plugins.
  3. Open Permissions.
  4. Choose the default you want. "Always ask" is the most careful.

Connected app lists and how to disconnect

ChatGPT (Connecting and managing app accounts in ChatGPT (opens in a new tab)):

  1. Open Settings, then Plugins.
  2. Select the plugin or app.
  3. Open the account's more options menu (•••), or use Connection.
  4. Select Disconnect.

Gemini (Use and manage Connected Apps in Gemini (opens in a new tab)):

  1. Go to gemini.google.com.
  2. In the sidebar, select Customize (select More first if you don't see it).
  3. At the top, select Connectors.
  4. Find the app, select More options, then Disconnect.

Your Google Account's links with other apps (Manage links between your Google Account and apps from other developers (opens in a new tab)). Start on your Google Account's linked apps page, then:

  1. Choose the kind of link: Sign in with Google, Linked account, Access to, or the Agent access filter.
  2. Select the app and select See details.
  3. Choose Stop using Sign in with Google, Delete link, Remove access, or Stop using for an agent.
  4. Confirm.

Claude (Use connectors to extend Claude's capabilities (opens in a new tab)):

  1. Go to Customize, then Connectors.
  2. Find the service in the Connectors list.
  3. Disconnect it.

What disconnecting does and doesn't delete

  • OpenAI: stops future access. Does not automatically delete existing or archived conversations, saved files or memories. Deleting a chat removes the connected-app data held in that chat, but never the original data in the other app. Deleting and archiving chats in ChatGPT (opens in a new tab)
  • Google: removing access stops the app reaching your account. Removing a sign-in link does not delete data held on the other app. Removing an agent's access does not remove the link itself.
  • Anthropic: as of 2 November 2026, its page still says nothing about deletion.

The book's reminder: reconnecting may not restore everything exactly, so write down how to reconnect before you disconnect.

Apps from other makers

Google says it "does not monitor or secure data from custom third-party Connected Apps" (Gemini Apps Privacy Hub (opens in a new tab), updated 26 October 2026).

For the "Permissions check" exercise, the steps above show where each connected-app list sits, as of 2 November 2026.

From the book

In the glossary wording, a harness (or agent harness) is:
The software around a model that turns it into a working agent. It holds the instructions, connects the tools, applies the permissions and keeps a log of each step.
Your First Rung, chapter 16

Sources

  1. Introducing the Model Context Protocol (opens in a new tab) · Anthropic, 2024
  2. Get started with Claude in Chrome (opens in a new tab) · Anthropic, 2026
  3. Use connectors to extend Claude's capabilities (opens in a new tab) · Anthropic, 2026
  4. Gemini Apps Privacy Hub (opens in a new tab) · Google, 2026
  5. Manage links between your Google Account & apps from other developers (opens in a new tab) · Google, 2026
  6. Use & manage Connected Apps in Gemini (opens in a new tab) · Google, 2026
  7. ICO tech futures: Agentic AI (opens in a new tab) · Information Commissioner's Office, 2026
  8. Linux Foundation Announces the Formation of the Agentic AI Foundation (AAIF), Anchored by New Project Contributions Including Model Context Protocol (MCP), goose and AGENTS.md (opens in a new tab) · The Linux Foundation, 2025
  9. Architecture overview (opens in a new tab) · Model Context Protocol, 2026
  10. Governance and Stewardship (opens in a new tab) · Model Context Protocol, 2026
  11. Security Best Practices (opens in a new tab) · Model Context Protocol, 2026
  12. Specification (opens in a new tab) · Model Context Protocol, 2026
  13. Tools (opens in a new tab) · Model Context Protocol, 2026
  14. What is the Model Context Protocol (MCP)? (opens in a new tab) · Model Context Protocol, 2026
  15. Connecting and managing app accounts in ChatGPT (opens in a new tab) · OpenAI, 2026
  16. Managing app permissions in ChatGPT (opens in a new tab) · OpenAI, 2026
  17. MCP joins the Agentic AI Foundation (opens in a new tab) · Soria Parra, David, 2025