Back to blog

How to choose a trustworthy MCP server

Seven checks before you connect an MCP server: publisher, source code, tool list, authentication, where it runs, maintenance, and what "verified" really means.

SOSourceWeftNov 20, 20266 min read

Last updated: 2026-11-20

The short answer

Trust an MCP server the way you would trust a browser extension: by who made it and what it is allowed to do. Before you connect one, check seven things: the publisher, the source code, the tool list, the authentication method, where it runs, how well it is maintained, and what any "verified" label actually covers.

No directory badge can tell you that a server's tools are safe. A badge can tell you who published it.

Why it matters

An MCP server receives tool arguments from your conversation and can act on external systems with your credentials. That creates three kinds of risk:

  • Over-broad access: the server can do more than the task needs, such as delete when you only wanted read.
  • Data exposure: whatever the model sends as a tool argument reaches the server's operator.
  • Prompt injection: text returned by a tool can contain instructions aimed at the model. A server that reads web pages, issues, or email passes that text straight into your conversation.

The checks below reduce these risks. They do not remove them.

The MCP project's own guide, Local Server Security, covers the same ground in more depth.

The checklist

1. Who publishes it?

Prefer a server published by the service it connects to. A Stripe server from Stripe carries less risk than one from an unknown account with the same name.

Look for a namespace that maps to a domain or organization you recognize, and for a link from the service's own documentation to the server.

2. Can you read the source?

An open repository lets you, or someone you trust, see what the tools do. Check that the repository linked from the listing is the one the package is built from, and look at its history: several contributors and a visible issue tracker are better signs than a single commit.

A closed-source remote server is not automatically bad. It means you are trusting the operator, so weigh the publisher check more heavily.

3. What tools does it expose?

Read the tool list before you connect. For each tool, ask whether it reads or writes, and whether you need it.

  • A server with five read-only tools is a small surface.
  • A server with delete_*, send_*, or execute_* tools deserves a closer look.
  • A server that publishes no tool list at all, for example because it requires sign-in first, cannot be assessed up front. Connect it with a low-privilege account and read the list then.

4. How does it authenticate?

From better to worse:

  1. Per-user OAuth sign-in: you grant specific scopes and can revoke them.
  2. A scoped API key: acceptable if the key is limited to what the task needs.
  3. A full-access key or a shared long-lived token: avoid where you can.

Whatever the method, grant the narrowest scope the service offers.

5. Where does it run?

  • Remote (Streamable HTTP): the server runs on the operator's infrastructure. It sees what you send, and nothing else on your machine.
  • Local (stdio): the server runs as a process under your user account. It can touch whatever you can, so the package you install matters. Pin a version instead of running "latest", and prefer packages from a known publisher.

Neither is safer in general. Remote shifts trust to the operator and local shifts it to the code.

6. Is it maintained?

Check the last update date and the version history. The protocol itself changes: the current MCP revision, 2026-07-28, removed the connection handshake, so a server that has not been touched in a year may not work with newer clients.

A recent release, a changelog, and answered issues are good signs.

7. What does "verified" mean here?

Read the fine print on any trust label. The word means different things in different places. On SourceWeft and in the official MCP Registry, "verified" means the publisher's namespace was checked, for example that it belongs to a domain or a GitHub account. It does not mean anyone reviewed the code or tested the tools. Some directories use the label differently, including as part of a paid listing.

Treat "verified" as an answer to check 1 at most, not as a substitute for checks 3 to 5.

Red flags

  • The name imitates a well-known service, but the publisher is unrelated.
  • The repository link is missing, dead, or points somewhere else.
  • Installation asks for a full-access key when a scoped one exists.
  • The tool list is far broader than the stated purpose.
  • The README tells you to disable confirmations or approval prompts.
  • The local install command pipes a remote script into a shell.

Where to find this information

A good directory puts these facts on the server's page so you do not have to dig for them. On the SourceWeft MCP directory, each server page shows:

CheckWhere it is on the page
PublisherPublisher line and the Trust field, with a note on what was verified
SourceRepository and Homepage links
ToolsThe Tools tab, or a notice when the server needs sign-in before its tools can be read
AuthenticationThe Auth field
Where it runsThe Transport and Runtime fields
MaintenanceThe Updated date and the Versions tab

Each page also carries a security note stating what the directory's checks do and do not cover.

After you connect

  • Start with a test account or a sandbox workspace.
  • Keep approval prompts on for tools that write, send, or delete.
  • Enable the server only in the conversations that need it.
  • Review the granted scopes after the first week and remove what you did not use.
  • Revoke access when you stop using the server.

FAQ

Are unverified MCP servers safe? Not by default. An unverified server may be fine, but nobody has confirmed who publishes it. Review the publisher, the tools, and the credentials it asks for before enabling it.

Does "verified" mean the server is safe? No. Where it means anything, it means the publisher's identity was confirmed. It says nothing about what the tools do, and its meaning varies between directories.

Is a local MCP server safer than a remote one? Not in general. A local server keeps your data off a third party's infrastructure but runs code with your user permissions. A remote server cannot touch your machine but sees everything you send it.

Should I prefer official servers? Yes, when one exists. The service's own server is maintained by the people who run the API and is the least likely to be abandoned or impersonated.

Can an MCP server read my whole conversation? Not by design. The MCP specification says servers "should not be able to read the whole conversation". The full history stays with your AI application, and a server gets only what the application sends it, which is normally the arguments of the tool calls made to it. That is a rule the application is expected to enforce, not a guarantee from the protocol. The model writes those arguments, so anything in the conversation can end up in one. Do not paste secrets into a chat that has servers connected.

How do I remove a server's access? Disconnect it in your AI application, then revoke the token or OAuth grant in the service's own account settings. Doing only the first leaves the credential valid.

Shareable URL

Keep reading

Related essays

All posts

How to install skills and MCP servers in Claude Code, from an empty folder

Start from an empty folder and install a skill and an MCP server for Claude Code: search, inspect before installing, where the config files go, and how to check and remove them. Every command was actually run.

Read next

MCP servers by category: the 17 categories in our directory and how to browse them

A list of MCP servers sorted into 17 categories, from databases to browser automation to communication, with the number of servers in each as of 21 November 2026, and how to filter, search and install from a category page.

Read next

How to check an agent skill before you install it: what scan flags tell you, and what they don't

A skill is instructions plus, sometimes, scripts that run on your machine. Here is how to read a skill's scan flags, what each one means, and the checks a scan cannot do for you.

Read next