By James Wilson · Jul 7, 2026

MCP Servers Every Developer Should Know About

The first time you add an MCP server to your AI assistant, something clicks. The assistant stops being a very good text generator and becomes something that can actually check, fetch, and do. The second thing that happens: you start wondering what else you can connect.

I have been running an MCP-heavy setup for months now, across Claude Code and Claude Desktop. This is the tour I wish someone had given me at the start: the categories of MCP servers that earn their place in the config file, what each unlocks, and how they combine. Not a directory dump of two hundred servers; a short list of the ones that changed how I work.

First, a 60-Second Refresher on MCP

The Model Context Protocol is an open standard, introduced by Anthropic, for connecting AI assistants to external tools. An MCP server is a small program (usually started by your AI client as a local subprocess) that exposes a set of typed tools: "list X", "create Y", "query Z". The assistant sees the tool descriptions, decides when to call them, and works with real data instead of whatever you pasted into the chat.

One config entry per server. Most clients (Claude Desktop, Claude Code, Cursor, Windsurf, Zed, and the rest) use the same JSON shape: a command, args, and environment variables for credentials.

Category 1: Code and Filesystem

The reference servers maintained in the official MCP ecosystem cover the basics every developer needs:

Filesystem: read, write, and search files in allowed directories. This is the backbone of "look at my project and fix it" workflows in clients that do not ship their own file access.
Git / GitHub: repository operations, issues, and pull requests. The practical win is triage: "summarize the open issues labeled bug and tell me which ones look like duplicates" is a one-liner instead of an afternoon.

If you use Claude Code, file and git access are built in, so you may not need these. In Claude Desktop or other chat-first clients, they are essential.

Category 2: Databases

Database MCP servers (Postgres is the most established) let the assistant inspect schemas and run queries. Two honest notes from experience:

For exploration, this is superb. "What tables reference users and which ones have grown the most rows?" turns schema archaeology into conversation.
For production databases, connect with a read-only database user. Not because the assistant is malicious, but because you would apply the same rule to any new tool with query access. Least privilege is free.

Category 3: Browser Automation

Servers like the Playwright MCP server give the assistant a real browser: navigate, click, fill forms, take screenshots. This is the one I underestimated most. Use cases that stuck:

"Open the staging site, walk through the signup flow, and tell me where it breaks" - a smoke test without writing a test.
Reproducing a user-reported bug while you keep working: the assistant drives the browser and reports back with screenshots.

Category 4: Your Infrastructure's Health

Here is the gap in most MCP setups I see: the assistant can read your code, query your database, and drive a browser, but it has no idea whether your services are actually up. Ops context is missing.

That is the slot the UptyBots MCP server fills. It connects the assistant to your uptime monitoring: 15 tools covering monitor listing and details, incident history, hourly and daily performance stats, notification history, monitor creation for all six check types (HTTP, API, ping, port, SSL, domain expiry), and pause/resume/delete.

Config is the same shape as everything else in this post:

{
  "mcpServers": {
    "uptybots": {
      "command": "npx",
      "args": ["uptybots-mcp-server"],
      "env": {
        "UPTYBOTS_API_KEY": "upty_your_key_here"
      }
    }
  }
}

Why it matters in combination with the other categories is the real point, so let me show that.

Where It Gets Interesting: Servers Working Together

A single MCP server is a convenience. Several together are a different way of working, because the assistant can correlate across them in one conversation.

Deploy with eyes open. In Claude Code: the assistant fixes a bug, runs the tests, deploys - and then, through the monitoring server, watches the HTTP monitor for the next check cycle and confirms production is healthy. Code context and ops context in the same session.

Incident with a suspect list. An alert fires. You ask: "The checkout monitor went down 20 minutes ago. Check the incident details, then look at the recent commits and tell me if anything deployed in that window touched the checkout path." Monitoring server provides the timeline, git server provides the suspects. (For the full version of this workflow, see AI Agent Incident Response.)

Onboarding a project end to end. "Clone the client's repo, find every external URL and service it depends on, and create uptime monitors for each." Filesystem server reads the code, monitoring server creates the checks. New project, instrumented in minutes.

Practical Advice for Your Config File

Add servers one at a time. Each server adds tools to the assistant's context. A dozen half-used servers make tool selection worse, not better. Start with what you will use this week.
One credential per server per machine. Named API keys and tokens (like upty_... keys named "claude-laptop") make revocation surgical instead of traumatic.
Prefer npx/official packages over random forks. An MCP server runs on your machine with your credentials in its environment. Read the source or stick to servers whose source you can read - the UptyBots one is open on GitHub (uptybots/mcp-server).
Ask the assistant what it can do. After adding a server, "what tools do you have from uptybots?" is the fastest sanity check that the config landed.

Frequently Asked Questions

Do MCP servers work with clients other than Claude?

Yes. MCP is an open standard. Cursor, Windsurf, Cline, Continue, Zed, VS Code with Copilot, JetBrains AI Assistant, and Amazon Q Developer all support MCP servers, including everything mentioned in this post.

Are MCP servers safe to run?

They are local processes with the credentials you give them. Apply the usual judgment: use official or open-source servers, scope credentials to the minimum needed, and keep one credential per machine so revocation is easy.

What does the UptyBots MCP server cost?

The server itself is free and open source. It needs a UptyBots API key; accounts start free with $5 in credits, and paid plans start at $4.99/month. Read operations work even with an empty balance.

How do I set up the monitoring server specifically?

Full setup for every client is in the MCP documentation, and a hands-on terminal walkthrough is in the Claude Code guide.