Find a connector or skill and check it's safe before you add it, in 15 minutes

The problem

It's Tuesday night, the kids are finally down, and you're scrolling a list of tools that promise Claude will keep track of who still owes you.
One has a checkmark, one says Community, and one is a link somebody dropped in a Facebook group with "this changed everything."
You can't tell which of them gets to see your client list, so you either add all of them or you close the laptop and add nada.

Steps

1. Start in Anthropic's directory, not in the group chat.
In Claude, click Customize in the left sidebar. You'll see three tabs: Skills, Connectors and Plugins. Inside each one, Discover shows what you can add. That's Anthropic's official directory, and it's the same catalog whether you're on the web, the desktop app or your phone. Start here because anything listed has to follow Anthropic's Software Directory Policy: it needs a privacy policy that explains what it collects and how long it keeps it, it's only supposed to collect what it needs to do its job, the developer has to give real contact and support info, and Anthropic reviews listings when they go in and keeps reviewing them after. That's a floor, not a guarantee, which is why the next steps exist.

2. Read the label before you read the description.
Click Connectors, then Discover, search for your app and open its page. Every connector carries one of these:

• Verified (a checkmark). Anthropic tested its tools for quality and compatibility, and it met the directory policy when it was reviewed. Anthropic says plainly this isn't a security audit, and the developer can change the tools after the review.
• Community. A third-party developer built it. Anthropic screened it before listing it but didn't review it in depth, and Claude shows you a reminder about that before you connect.
• Custom. You added it yourself from a link. Anthropic hasn't reviewed it at all.

Here's the part people miss: once it's connected, a Community connector has the same access as any other connector you approve. The label tells you how closely Anthropic looked. It doesn't limit what the connector can do.

3. Check who made it.
On the connector's page, look at the developer details and the privacy policy. The question is simple: is this the company that runs the app, or somebody else? If a connector for your bookkeeping app was built by someone who isn't that company, find out who they are before you hand over your sign-in. Anthropic doesn't run third-party connectors' servers and doesn't control how they handle your data, so the developer's privacy policy is the one that counts. And about popularity: the directory ranks connectors by usage, like an app store. A lot of people using something tells you it's popular. It doesn't tell you anybody checked it.

4. Match what it can do to the job you actually have.
The connector's page shows what it can read and what it can write. A connector can be built to read your data, create, change or delete things, and act for you. Write your job down in one line first. "Tell me who has a session Thursday" is a reading job. If the connector also sends email or deletes events, ask yourself why you'd need that. Most of us, at the start, only need Claude to look.

5. Read the sign-in screen like it's a contract.
When you click to connect, Claude sends you to the app's own sign-in page. You type your password there, never in the chat. That page lists what access you're approving. Anthropic's advice is to review those permissions carefully. If it asks for more than your one-line job, close the tab. Nothing is connected until you approve it.

6. Turn the dials down after you connect.
Go back to Customize, then Connectors, and select the connector under Your connectors. Under Tool permissions, you can set each tool (or a whole group of them) to Always allow, Needs approval or Blocked. Set anything that sends, changes or deletes to Needs approval or Blocked. In a chat, when Claude asks to use a tool, you'll see Allow once and Always allow; Anthropic says to click Always allow only for servers you trust. To keep a connector out of one conversation, click + in the message box, then Connectors, and switch it off. If you stop using it, Disconnect signs Claude out, and Remove (in the three-dot menu) takes it off your account.

7. Skills and plugins: same idea, different check.
A skill is written instructions, sometimes with scripts, that Claude follows when your task matches it. Under Customize, Skills, your skills are grouped by where they came from, including From Anthropic & Partners. Anthropic says skills that someone shares with you, or that you upload yourself, aren't reviewed by Anthropic, so open the skill and read its SKILL.md before you turn it on. The two risks they name are hidden instructions that steer Claude somewhere you didn't ask (prompt injection) and your information getting sent somewhere it shouldn't go (data exfiltration). Plugins bundle skills and connectors into one install. The ones in the directory get an automated check and a security scan on every version, plus a person reviews each new listing before it goes live. A plugin you add from an outside link or upload yourself doesn't get that review. And directory plugins update on their own with nothing for you to accept, so if one starts acting different, go look.