How to check an MCP server before you install it
AI assistants are getting hands. Through the Model Context Protocol (MCP), an AI assistant can call tools: read files, search a database, open a ticket, send a message. Each of those tools is provided by an MCP server, a small program that you add to your assistant. There are a fast-growing number of them, written by companies, open-source projects and individual developers.
That is useful, and it is also a supply-chain decision. An MCP server you install locally runs as a program on your computer with your permissions. A remote one sees whatever your assistant sends it. Before you add one, it is worth five minutes of checking. Here is what we look at, in order.
1. Who actually publishes it?
The name in a directory listing is a claim, not proof. Look for a link between the server and an identity you can check:
- Does the package (on npm, PyPI or a container registry) point to a source repository?
- Does that repository belong to the same person or organization as the package name?
- If the publisher says they are a company, does the company's own website link back to the server?
A server whose package, repository and website all point at each other is far easier to trust than one that exists only as a package with no source.
2. Can you see the code that you are running?
Open source is not automatically safe, but it can be checked. Things to look for:
- The repository is public and recently active.
- It has a recognized license, so you know what you are allowed to do with it.
- The published package was built from that repository. npm and PyPI both support provenance records (build attestations) that tie a package version to the source and build that produced it. When one exists, it is strong evidence that the code you read is the code you install.
3. What happens when you install it?
Some packages run their own scripts during installation, before you ever use them. That is not always bad (some need to compile something), but it means code runs on your machine at install time. Check whether the package declares install scripts, and read them if it does.
4. What changed in this version?
Most supply-chain incidents arrive in an update, not in the first version. Before you accept a new version, compare it with the one you have:
- Does it contact a new network address?
- Does it ask for new permissions or new environment variables (API keys)?
- Did the maintainer or the repository change?
A change log is helpful, but the package contents are what count.
5. Are there known security advisories?
Search the public advisory databases (for example the GitHub Advisory Database and OSV) for the package name. A past advisory that was fixed quickly is a good sign; an open one is a reason to wait.
6. What does the server ask your assistant to do?
MCP servers describe their tools to the assistant in plain text. That text is read by the model, so treat it like any other input you did not write. A tool description that tells the assistant to ignore your instructions, to send data somewhere, or to call other tools is a red flag.
7. Give it the least you can
Even a server you trust should get only what it needs:
- Use a separate API key with the smallest scope, not your main one.
- Prefer read-only tools when you only need to read.
- Remove servers you no longer use.
No warnings is not a guarantee
Every check above reduces risk; none removes it. A clean result means nothing known is wrong, not that nothing is wrong. Keep the evidence you looked at, and look again when the server updates.
Doing this at scale
We built UND Assay (undassay.com) to run these checks and show the public evidence behind each one, because doing them by hand for every server and every update does not scale. Whether you use a tool or a checklist, the habit is what protects you: know who wrote it, see what it is, and check what changed.