# Is this MCP server safe? A 10-point check before you install

Level 3: Connect AI to your tools · 8 min · Updated 19 November 2026 · Path: Connect MCP servers to your assistant

By the end you'll decide in a few minutes whether an MCP server is worth trusting, using its listing page and official sources.

You found an MCP server that does exactly what you need. Before you hand it your calendar, your code or your company's CRM, spend five minutes on this list. None of it needs technical skills; most answers are on the server's listing page in this directory.

Why bother? Anthropic's help center says custom connectors aren't verified by Anthropic and that a malicious server can hide instructions in what it returns. The MCP Registry's own docs say it leaves security scanning to package registries and downstream directories. Someone has to check, and for the servers you add, that's you.

## The ten points

### 1. Who publishes it?

Prefer the company behind the service ("the Linear server, by Linear") over a third party's wrapper. Check the website and repository links on the listing.

### 2. Is its name verified in the MCP Registry?

Registry names look like `io.github.acme/server` or `com.acme/server`. The registry only lets the owner of that GitHub account or domain publish under it. A name that matches the company's own domain is a good sign. The listing page shows the registry name when there is one.

### 3. Does it answer? (hosted servers)

For hosted servers, the listing page shows our recent checks: whether it answered, asked for sign-in or timed out. A server that's often down isn't worth the risk of half-finished actions.

### 4. Does it ask for sign-in, and how?

OAuth sign-in through the service's own login page is the normal pattern for servers that reach your account. Be wary of a server that asks you to paste a password, or a key with more access than the job needs. See [Why does this server ask me to sign in?](https://ai-powered.com/academy/connect-mcp-servers/why-sign-in)

### 5. Which tools does it offer?

Read the tool list on the listing. A calendar server that also has `run_shell_command` is a red flag. Fewer, specific tools are easier to trust.

### 6. Are tools marked read-only or destructive?

Servers can label tools with [annotations](https://ai-powered.com/academy/glossary/tool-annotations) such as `readOnlyHint` and `destructiveHint`. They help, but the MCP project is clear that they're hints, not guarantees: a dishonest server can mislabel a tool, and clients should treat annotations from untrusted servers as untrusted.

### 7. For a local server: what command does it run?

A local server runs a command on your computer with your permissions. Read the whole command. Expect something like `npx -y @company/server`; be suspicious of anything chaining extra commands, downloading scripts or touching folders it doesn't need.

### 8. Is it maintained?

Check the repository's recent activity and open issues on the listing. Abandoned servers don't get security fixes.

### 9. What does it read from outside?

Servers that fetch web pages, emails or tickets can pass along [prompt injection](https://ai-powered.com/academy/glossary/prompt-injection). That doesn't rule them out, but keep approvals on for any tool that writes.

### 10. Can you give it less?

Smaller scopes, a read-only database user, one folder instead of your home directory, a test account first. The less it can reach, the less a mistake costs.

*Diagram: Adding a server, step by step.* Pick a server you trust, add it to your app (a URL or a command), sign in or restart, then try a small task and approve the first tool call.

**Try it**

Pick a server you're considering and open its listing page here. Go down the ten points and note the answers. If three or more are "don't know", choose another server.

## Key takeaways

- Publisher, registry name and our checks tell you who you're trusting.
- Tools and annotations tell you what it can do; annotations are hints only.
- Give every server the least access that works.

## What to try next

- Practise on a read-only server: [connect ours in two minutes](https://ai-powered.com/academy/connect-mcp-servers/connect-ai-powered-mcp-server).
- Building your own? The [MCP server security checklist](https://ai-powered.com/academy/build-mcp-servers/mcp-server-security-checklist) is the builder's side of this list.

## Sources

Official documentation, checked on 19 November 2026:

- [The MCP Registry (modelcontextprotocol.io)](https://modelcontextprotocol.io/registry/about)
- [Tool annotations as risk vocabulary (MCP blog, 16 March 2026)](https://blog.modelcontextprotocol.io/posts/2026-03-16-tool-annotations/)
- [Security best practices (modelcontextprotocol.io)](https://modelcontextprotocol.io/docs/tutorials/security/security_best_practices)
- [Getting started with custom connectors using remote MCP (Claude Help Center)](https://support.claude.com/en/articles/11175166-getting-started-with-custom-connectors-using-remote-mcp)

Next lesson: [Why does this MCP server ask me to sign in?](https://ai-powered.com/academy/connect-mcp-servers/why-sign-in)
