A directory makes discovery easy and judgment hard: every entry looks equally plausible in a table. Before you hand a server credentials, five checks predict most of what will go wrong.
1. Is anyone maintaining it
Look at the last commit and the open issue count together. Old code with no issues is often finished; old code with twelve unanswered issues is abandoned. Neither number alone tells you which.
The useful question: is there a release in the last six months, or a response to an issue? A maintained-but-quiet project is fine. A project whose author stopped replying is a dependency you inherit.
2. Which transport, and does that match your client
A stdio server is launched by the client as a subprocess; an HTTP server is a service you run. Most desktop clients only launch stdio servers, which is why an HTTP-only server with great documentation still cannot be installed by clicking.
3. What credentials does it need, and how wide are they
A server that wants a personal access token with full scope is asking for more than the task usually requires. Prefer tokens scoped to one repository, one calendar, one database role. If the documentation only shows the maximal token, that is itself information.
4. How much can it destroy
This is the check people skip. Ask: if the agent misreads a request, what is the worst outcome? Reading a database — annoying. Writing to production — an incident. Deleting files — an incident with no undo.
The rule we apply in our own entries: anything irreversible gets a warning in the entry, not just in the README.
5. Does the install command still resolve
Run it before you configure anything. A surprising number of listed packages are renamed, unpublished, or point at a moved repository. Running `npx -y <package> --help` takes ten seconds and settles it.
What the tiers mean
Our directory labels each entry **reviewed** or **listed**. Reviewed means a person did these five checks and wrote down what they found, with a date. Listed means the package exists and declares an executable — nothing more. If your choice is expensive, restrict yourself to reviewed entries and treat the rest as leads.
Next: debugging a server for when the install works and the tools do not.