Agents Index

MCP Server Security Checklist: 12 Checks Before You Connect

Updated 2026-11-18

Before you connect an MCP server, check who runs it, where it runs, what it can do, and what it can read. A local server runs code with your permissions. A remote server sees your requests. Any server can place text in front of the model through tool descriptions and results, so connect only servers you trust and give them the least access that works.

How MCP servers can go wrong

An MCP client puts a server's tool names, descriptions and results into the model's context. The model treats that text as input. That is the basis of most MCP-specific attacks:

Both Anthropic and OpenAI warn about this in their client docs. Claude Code tells users to verify they trust each server and notes that servers that fetch external content can expose you to prompt injection. OpenAI states that custom MCP servers are not developed or verified by OpenAI.

Checklist

1. Decide whether you trust the publisher

Look at who runs it. Servers in the official MCP Registry use namespaces tied to a GitHub account or a domain, which shows who registered the name. It does not show that the code is safe. The registry delegates security scanning to package registries and downstream aggregators. Check the repository, the maintainers and recent activity.

2. Prefer remote servers from known vendors when you can

A remote server runs on the provider's infrastructure, so it cannot read your files or run commands on your computer. The trade-off is that it receives whatever the model sends. A local server runs on your machine and has direct access to your system. Neither is automatically safer. Pick based on what data the server will see.

3. Sandbox local servers

The MCP security guidance says local servers can run arbitrary commands with the privileges of the client. It recommends running them in a sandbox with minimal default privileges and restricted file system and network access. Use a container or a dedicated low-privilege user for anything you did not write. Read the exact command and args before you approve a config, especially a one-click install or a repository's .mcp.json. Claude Code asks for approval before it uses project-scoped servers for this reason. VS Code applies Workspace Trust to workspace MCP files.

4. Pin versions

If a local server is installed with npx or uvx and no version, you run whatever the latest release is on every start. Pin an exact version:

claude mcp add --transport stdio example -- npx -y example-mcp-server@1.2.3

Update on purpose, after reading the changelog. Docker images are safer pinned by digest than by latest. For remote servers you cannot pin the code, so re-read the tool list when the provider announces changes.

5. Review the tool list before you connect

Read every tool name, description and parameter. Look for tools that write, delete, send or spend, and for descriptions that give the model instructions unrelated to the tool. The MCP specification says there should always be a human in the loop with the ability to deny tool invocations, and clients should show which tools are exposed to the model. Disable tools you do not need. ChatGPT lets you turn individual tools on or off on the app's settings page, and Claude has a "Search and tools" menu for the same purpose.

6. Re-check after updates

A server can change its tools over time. Servers can announce a changed list, and ChatGPT lets you refresh an app to pull updated tools and descriptions. Treat a changed tool list as a new review, not a routine update. Anthropic's connector guidance tells users to monitor tool behavior changes.

7. Use least privilege

Give the server only what the task needs: a read-only API token instead of an admin key, one repository instead of an organization, a test account instead of production. On Team and Enterprise plans, a shared API key means everyone who uses a Claude connector reaches the service with the same credential, so Anthropic advises using a limited one.

8. Prefer read-only tools

If you build a server, split reading and writing into separate tools and mark read-only tools with the readOnlyHint annotation. ChatGPT respects it: write actions require confirmation by default, and tools without the hint are treated as write actions. One caution from the MCP specification: clients must treat tool annotations as untrusted unless they come from a trusted server. A hint is not a guarantee. If you consume a server, check that tools labeled read-only really are.

9. Keep approvals on for write actions

Review the tool input before you approve. OpenAI notes that an incorrect write can destroy, alter or share data, and lets you remember an approval for the rest of a conversation. Do that only for tools you trust. Anthropic notes that Research can call connector tools without further approval, so disable write-capable tools before you use it.

10. Check OAuth scopes

When a server uses OAuth, read the consent screen. The MCP security guidance warns about omnibus scopes such as files:*, db:* and admin:*, because a stolen token with broad scopes opens everything. If you build a server, start with a minimal scope set of low-risk read operations and ask for more when a privileged tool is first used. Avoid publishing every scope in scopes_supported and avoid wildcard scopes.

11. Build servers that validate everything

Server authors have their own list. The MCP specification says servers must validate all tool inputs, implement access controls, rate limit tool calls and sanitize outputs. Servers must also not accept tokens that were not issued for them. This is the "token passthrough" anti-pattern, which the spec forbids. If your server proxies a third-party API with a static client ID, implement per-client consent so one user's earlier approval cannot be reused by an attacker.

12. Prefer endpoints that are alive and accountable

An unreachable endpoint cannot be reviewed, and an abandoned one can change hands. According to Agents Index probe data, Nov 2026, of the registry's remote endpoints we probed, 23,704 answered without sign-in, 13,608 required sign-in and 7,189 were unreachable. A server that answers without sign-in can be called by anyone, so check what data it exposes. See the State of MCP report and how we probe.

Quick pre-connect routine

  1. Find the server on a registry or directory, such as the official registry listing, and open its repository.
  2. Check the auth type and the scopes it asks for.
  3. Connect with a throwaway or read-only credential first.
  4. List the tools and read them.
  5. Turn off every tool you will not use.
  6. Pin the version if it is local.

Setup guides: Claude Code, ChatGPT. Browse popular servers to see what others connect.

Sources

FAQ

Are MCP servers safe to use?

It depends on the server. A local server runs code with your permissions, and any server can feed text to the model. Connect only servers you trust, review the tool list, and keep write access to a minimum.

What is tool poisoning in MCP?

Tool poisoning is when a malicious or compromised server hides instructions in tool names, descriptions or results so the model acts on them. Review tool lists and treat annotations from untrusted servers as untrusted.

Is a remote or a local MCP server safer?

Neither is automatically safer. A local server can run arbitrary code on your machine. A remote server sends your data to another party. Sandbox local servers and give remote ones only the OAuth scopes they need.

How do I pin an MCP server version?

For npm servers, name an exact version, such as npx -y package@1.2.3, instead of latest. For remote servers you cannot pin code, so re-check the tool list after any change.

More guides