The tools and information Claude can reach

The last piece of setting AI up is what it can actually reach: your files, your systems, the other apps you already run. MCP servers are the standard way to plug those in, so the AI works with your real things instead of guessing at them.

Where you set one up decides who gets it. Put it in the project, in a file the repo carries, and everyone who works on that project gets it too. Put it in your own settings and it is yours alone. That distinction matters more than it looks: a server only in your settings works on your machine and quietly does not exist on anyone else’s, so a build that leans on it cannot be handed over.

Whatever the server needs to log in stays out of the shared file. The file names the key it needs and reads the actual value from your environment, so the setup can be committed and shared without the secret ever being in it.

Common questions

1. Which way round is worse to get wrong?
   A shared server left in your own settings. It still works for you while silently not existing for anybody else, which is worse than the other mistake because the work looks reproducible and is not.
2. And the other way?
   A personal server in the project file breaks for everyone else the moment they clone the repo. Noisy rather than silent, so at least you find out.
3. How do I share a server without sharing my credentials?
   The config file holds a reference to the variable and your environment holds the value, so the file can be committed complete without ever carrying the secret. A README pointing at a dummy env file does the rest.
4. Do configured servers only load when they are used?
   No. Every tool from every configured server is put in front of the model on every request, so they all consume tokens and they all compete for its attention while it works out what you want done.
5. So what does that cost me?
   Configuring a server is cheap, and it is not free. A smaller, sharper set of tools is easier for the AI to choose well from than a large one.
6. Can I keep a server without paying that cost?
   Two ways. Defer the tools behind a search tool, so their descriptions stay out of the context until it looks for them. Or allowlist a server's tools, turning them off by default and naming only the ones that should reach the model.
7. My custom tool keeps losing to a built-in. Why?
   A description that says only what your tool does loses to something like grep, because the model understands grep better. The line that wins is the boundary condition, when to use this one instead of the obvious alternative, not the capability.
8. Should I write my own server or take an existing one?
   For a standard integration like an issue tracker or a code host, take the community server, because someone else maintains it. Write your own only for a workflow that is yours alone, which is the part actually worth owning.
9. Is a server only a set of tools?
   No, two things get published: the tools, and a resource listing of what the server holds. Publishing the resources means the agent does not spend its opening calls on discovery, it reads what exists and then acts.

Back to Ways to set AI up to help. The craft around it is Working with AI.