MCP, plugins, and tool trust

Home Tools Platforms Claude Code MCP, plugins, and tool trust

Add narrowly scoped external capabilities and inspect the extension before loading it.

Coding & development · Documented · Documentation checked 2026-10-04

Claude Code in terminals and supported development surfaces; Claude chat is not a local shell.

Authentication, installed version, surface, subscription or API billing, and organization policy matter.

The worked example is illustrative and has not been tested in a signed-in product account.

What it does and how it connects

Add narrowly scoped external capabilities and inspect the extension before loading it.

MCP exposes external tools and context; plugins bundle extension capabilities such as skills, connectors, and supported code features. Installing a bundle can add several kinds of authority. A tool description is an interface contract, not proof that its server is trustworthy.

Walkthrough

Use the product named in the access panel. The steps below combine documented controls with the editorial example.

  1. Inspect the extension’s publisher, files, tool operations, and required credentials.
  2. Connect a trusted server or install the package through the documented client path.
  3. Authorize only a sample account or read-only resource needed for the task.
  4. Inspect tool responses and permission behavior before enabling writes.

Worked example

A coding agent might read one approved issue from an external tracker, then edit a local branch. It does not need access to every project or permission to close issues. Treat the tracker connection as a separate component of the task.

Read only [approved issue] using the connected tracker. Confirm the project and issue identity, then propose a local implementation plan. Do not modify the tracker, post comments, change status, or access unrelated projects. Stop if the correct issue cannot be verified.

Inspect and verify the result

Check tool activity and returned issue identity. Review the server’s actual account scopes. A successful read does not establish that a future write is appropriate.

Limits and troubleshooting

Remote connectors need reachable services; local servers depend on host access. Plugin synchronization can require the supported account sign-in and client version. An unavailable integration should be reported rather than replaced with guessed context.

Data, permissions, and actions

External servers receive tool requests and may store data or perform actions. Local tools can reach host files. Inspect authentication, returned instructions, and logs. Never assume plugin installation creates safe permissions automatically.

Subagents and hooks

Separate delegated reasoning from lifecycle-triggered automation and shared side effects.

Explore Subagents and hooks

A local change with explicit permission review

Implement a small sample fix and verify its behavior.

Explore A local change with explicit permission review

A read-only extension-assisted review

Use approved issue context and delegated review without external side effects.

Explore A read-only extension-assisted review

Plugins, connectors, and skills

Separate packaged capabilities, reusable instructions, and the connections that let AI reach other software.

Explore Plugins, connectors, and skills

Sources and verification

Documentation checked 2026-10-04. Examples are illustrative unless described otherwise.

Features and extension relationships · Anthropic

Use plugins in Claude · Anthropic

Share this page