Skip to content
coworkplugins
Guide

Are Claude plugins safe? What gets checked, what doesn't, and what to look for

What Anthropic reviews, what this directory checks before listing a plugin, what nobody checks, and what to look at before you install one.

By Arun BabuLast updated Beginner

A Claude plugin can be as harmless as a page of instructions or as powerful as a program that runs on your computer, reads your email and spends your ad budget. So the honest answer to "are Claude plugins safe?" is: it depends on the plugin, and on who has checked it. This page explains what Anthropic checks, what this directory checks, what nobody checks, and what to look at yourself.

What a plugin can actually do

A plugin bundles parts that carry different amounts of risk:

  • Skills and commands are written instructions that Claude follows. On their own they run no code, though they can tell Claude to use tools it already has.
  • Connectors reach outside services, such as your CRM or email, with the account you connect. What they can read and change is set by the permissions you approve at sign-in.
  • Agents and hooks run commands on your computer when certain events happen, in Cowork and Claude Code.
  • Local MCP servers are programs that run on your computer.

Anthropic's Claude Code docs say it directly: a plugin you install can run code with your user privileges. A plugin made only of skills is a different thing from one with hooks and a local server, even when both are called plugins.

What Anthropic reviews

Anthropic reviews plugins listed in its own directory. According to its plugins overview, every version gets automated validation and a security scan, and a person reviews a new listing before it goes live. When we ran claude plugin validate on a test plugin, Claude Code itself noted that the directory "also runs a safety scan and a policy review that this command does not".

That review does not cover a plugin you add from a marketplace URL or upload yourself. Most community plugins, including most in this directory, are installed that way, so Anthropic hasn't reviewed them.

For connectors, Anthropic uses labels. Verified means tested for quality and compatibility, which is not a security audit. Community means built by a third party, screened but not reviewed in depth. Custom means you added it and Anthropic hasn't reviewed it.

What this directory checks before listing a plugin

Community plugins here are found by an automated search of public GitHub repositories. Before one is listed, it has to pass these checks:

  • a public repository that isn't a fork, an archive or a template, with a recognised licence;
  • updated in the last 60 days, and mentions Claude;
  • a plugin manifest at the repository root, so it is a real plugin rather than a loose skill or a server;
  • a README of at least 800 characters with install or usage steps;
  • at least three real skills, commands or agents (an MCP server alone doesn't count);
  • at least one GitHub star;
  • not a near-copy of another listing, and no mention of evading checks or bots.

Plugins that pass go to the site owner as a short summary to approve or decline. That approval judges whether it belongs in the directory. It is not a code review.

What nobody checks

Neither step reads a community plugin's code. Nobody here has checked what its hooks do, which servers it contacts, whether it collects analytics, or whether it works as described. Stars show interest, not safety. That part is up to you, and it doesn't take long.

What to look at before you install

When we compared plugins for memory, sales and marketing, every one of these came straight from a plugin's own README:

  • Does anything leave your machine? One popular memory plugin has anonymous usage analytics on by default. Another sends the whole conversation to a third-party service to decide what to keep.
  • Can it act for you? Some marketing plugins can publish posts, send email or change ad budgets. Most ask first, but one has an email mode that sends "without anyone reading them, within a daily cap".
  • What runs on its own? Hooks can run at every session start, prompt or compaction. One plugin writes new skills into your project in the background, and another continues Claude's turn after code edits.
  • What does it need? Some require an account with the vendor, a paid API key or a token stored on the vendor's servers.
  • What does it change? Some edit your Claude settings or your shell profile during setup.
  • What is the licence? Most are MIT or Apache-2.0, but some are source-available or AGPL.

None of this is necessarily bad. Analytics can be fine, and acting for you is the point of some plugins. What matters is knowing before you install, not after.

Anthropic's advice for third-party connectors

Anthropic's own guidance for any connector built by someone else is short:

  • Only connect to servers from developers and organizations you trust.
  • A connector's developer controls which tools it exposes and can change them at any time.
  • Review the permission scopes requested during sign-in.
  • Stay aware of prompt injection, even though Claude has built-in protections.
  • Watch for unexpected changes in how a tool behaves.

What each part of a plugin can reach, layer by layer, is in what a Claude plugin can access.

A five-minute check

  • Read the README's install and privacy sections. Look for the words telemetry, analytics, API key, account, send, publish and cloud.
  • Look at what's inside. On GitHub, open the hooks folder and the .mcp.json file if they exist: they show what runs on your computer and which servers it talks to. In Claude Code, claude plugin details <plugin> lists every skill, agent, hook and server after installing.
  • Check who made it and how active it is. Each plugin page here shows the author, licence, stars and last update, and links to the source.
  • Connect the minimum. Approve only the connector permissions the task needs.
  • Try it in a throwaway project first, with test data rather than your real CRM or ad account.
  • Know how to undo it. Turning a plugin off is safer than removing it while you test. Removing a whole marketplace also uninstalls every plugin from it. Details in turn off or remove a plugin.

On Team and Enterprise plans, your admin can limit which plugins and marketplaces you can add, and can turn off local MCP servers entirely. If a plugin is blocked at work, that is usually why.

Official plugins

Anthropic's own plugins, such as Sales, Marketing and Finance, are published openly on GitHub, so you can read every file. They lean on skills and connectors. The Sales plugin's README, for example, says changes a skill suggests on its own are shown to you first, and that instructions found inside emails or transcripts are treated as information, not commands. See the official plugins, and for more on the difference, official vs community plugins.

Sources

Read more