Three daily-use MCP servers for Claude Desktop, and how to set them up

A developer writing on dev.to has shared the three Model Context Protocol (MCP) servers they keep running day to day in Anthropic's Claude Desktop, together with the short routine for installing each one. Rather than another survey of experimental integrations, the post describes a working setup built around filesystem access, GitHub and PostgreSQL — and closes with a caution about how much of a machine the filesystem server can expose.

Filesystem: reading the project directly

The first server, @modelcontextprotocol/server-filesystem, is installed via mcp-hub install. According to the post, its main value is removing copy-paste from the workflow: you point the server at a project root and Claude reads the files itself. The author's illustrative prompt asks the model to read an authentication module and explain how its token validation works.

GitHub: pull requests inside the conversation

The second, @modelcontextprotocol/server-github, gives Claude the ability to check issues, read pull requests and browse repositories. The author says this is what they mainly reach for during code review. A typical request, per the post, is asking the model to look at the changes in the last few pull requests and produce a summary without the developer opening a browser.

PostgreSQL: queries without leaving the chat

The third server, @modelcontextprotocol/server-postgres, provides direct database access from within the conversation. Instead of switching to a database client, running a query and pasting results back, the developer simply asks Claude something like which users registered this week but have not yet logged in, and the model runs the query itself.

The setup itself

Each server takes roughly two minutes to add, the post claims. The steps are the same in every case: run mcp-hub install for the package, add an entry to claude_desktop_config., and restart Claude Desktop so it picks up the new server. The author notes that all three packages were verified on npm and points to the mcp-hub CLI as a way to discover additional servers beyond these three.

The warning that comes with filesystem access

The post's most practical caveat concerns scope. The filesystem server, the author warns, can read the entire disk if it is configured that way, so the paths you hand it deserve deliberate thought. Configuring it against a specific project root rather than a broad directory limits what the model can touch. The same logic applies in spirit to the other two servers: a GitHub token or an open database connection is real access to real systems, and the convenience of in-conversation queries comes paired with the responsibility of scoping that access.

Why it matters

Most coverage of MCP has focused on the protocol itself and the growing catalogue of servers built for it. A first-person account of a setup that survives daily use is more useful to developers deciding where to start, because it narrows the choice to three servers that cover the places code actually lives: local files, the repository host and the database. It also frames the real trade-off. Handing an AI assistant direct access to your filesystem, repositories and production data removes a great deal of manual copying, but the security boundary becomes the configuration file rather than the clipboard — and that file deserves as much attention as the prompts sent through it.