MCP Servers

The Claim

The Model Context Protocol lets you give Claude access to external tools and data sources through standardized servers. Think of it as plugins, but with a protocol spec instead of a proprietary API. You can spin up MCP servers for databases, file systems, APIs, and anything else you want Claude to interact with directly.

What I Tried

I set up three MCP servers: one for a Postgres database (Neon), one for filesystem access scoped to my project directory, and one for fetching live documentation from a few sites. The goal was to let Claude Code query my database schema, read project files, and reference up-to-date docs all within the same conversation.

The database server was the most useful. Being able to say "look at the users table and write me a migration that adds a preferences column with a sensible default" and having it actually inspect the live schema before generating SQL is a real workflow improvement. No more copy-pasting `\d users` output into a prompt.

The filesystem server worked but overlapped heavily with what Claude Code already does natively. The documentation fetcher was hit or miss depending on the site's structure.

What Surprised Me

The configuration is more manual than I expected. You're editing JSON config files, managing server processes, debugging connection issues when something doesn't start cleanly. There's no `mcp init` that just works. I spent a solid hour getting the Postgres server authenticated correctly, and the error messages weren't helpful.

What also surprised me is how good it is once it's running. The protocol itself is well-designed. Servers are composable, the tool definitions are clean, and Claude handles multi-tool workflows smoothly. The gap between "configured" and "not configured" is massive.

Who It's For

Developers already using Claude Code who want to extend what it can access. If you're comfortable writing config files and debugging server processes, the payoff is real. If you want something that works out of the box with zero setup, you'll bounce off this hard.

Verdict

Conditional. The concept is sound and the protocol is solid. When it works, it meaningfully expands what's possible in a single Claude session. But the setup friction and ecosystem maturity aren't there yet for a blanket recommendation. I'm keeping my database and docs servers active and checking back monthly on community servers. This will likely graduate to "kept" within six months.