How to do it
This applies to any assistant that supports MCP.
- Decide which servers you actually need. Better three servers you use than ten connected "because they're there." Each extra server is one more process and one more tool for the AI to choose among.
- Put them all in the file, with distinct names. The operational syntax (here, files and memory together):
{
"mcpServers": {
"documents": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/yourname/Documents"]
},
"memory": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-memory"]
}
}
}
Each block is a server, separated from the previous one by a comma. The name (documents, memory) is yours to choose and serves to tell them apart.
- Restart the assistant completely, so it loads all the servers.
- Check that they're all there. Ask the AI "what tools do you have available?": it should list the ones from every connected server. If one is missing, that's the one to check.
A concrete example
Marta is working on a project and wants the AI to read the project files, remember the decisions made, and query a small client database. She connects three servers: filesystem on the project folder, memory for the decisions, database for the clients, all in the same file with different names.
When she asks "check the project status in the files and tell me which clients in the database haven't received the proposal yet," the AI uses two servers together: it reads the files with one and queries the database with the other, combining the information. Marta doesn't have to say which tool to use: the assistant chooses on its own based on the question. Three different capabilities, a single conversation.
When it does NOT work (and how to fix it)
If some tools disappear
Many assistants have a ceiling on the number of tools active at once (Cursor around forty, others higher): past that threshold, some tools silently stop being seen. If you've connected many servers, disable the ones you don't need for that job and stay under the limit.
If two servers have tools with the same name
With many servers you can run into a name conflict between tools, and the AI gets confused about which to use. Give the servers clear, distinct names, and if the problem persists, keep active only the servers needed for the task at hand.
If the computer slows down
Each server is a program that consumes memory and processor. With many servers active at once, the machine can struggle. Connect only the ones you actually use and close the assistant when you're done, freeing up the resources.
A tip from someone who actually uses it
Group servers by type of work and activate only the group you need. For document work, filesystem and memory; for analysis, database and Excel. Keeping on only the three or four servers for the task at hand makes the AI more precise in its choice of tools and the computer lighter. The temptation to leave them all connected all the time is the most common cause of slowdowns and disappearing tools.
Frequently asked questions
How many servers can I connect?
Technically many, but the practical limit is your assistant's tool ceiling and your computer's resources. Under four or five servers you're fine almost everywhere; beyond that, start counting the total tools and keeping active only the necessary ones.
Does the AI figure out on its own which server to use?
Yes: it reads the tool descriptions of each server and chooses based on the question. It works better if the servers have clear, distinct purposes; if two do similar things, it may hesitate. Distinct names and roles help it choose right.
If I manage many servers, is there a way to simplify?
Yes: for those who use many (more than five) there are "virtual" servers or gateways, which gather multiple servers behind a single entry point. They're a more technical step up, useful in a company or for those who orchestrate many tools; for personal use they're usually unnecessary.