The MCP Servers Worth Installing in 2026

The Model Context Protocol (MCP) did for AI tools what USB did for hardware: one standard plug. Instead of every assistant reinventing integrations, an MCP server exposes a capability — your filesystem, a database, a browser — and any MCP-aware client (Claude Code, Cursor, and many others) can use it.

The flip side of that success is noise. There are now hundreds of servers, ranging from essential to barely functional. This is a short, opinionated list of the free, open-source ones that have earned a permanent spot in our setup — plus how to judge the ones that haven't.

The official reference servers

Start here. These are maintained by the MCP steering group in the modelcontextprotocol/servers repo, which means they track the spec closely and aren't going to disappear next quarter.

• Filesystem — scoped read/write access to directories you explicitly allow. The foundation for almost any coding workflow.
• Git — inspect history, diffs, and branches without the agent shelling out blindly.
• Fetch — pull a URL and hand the agent clean, readable content. The simplest way to ground answers in real pages.
• Memory — a knowledge-graph store so an agent can remember facts across sessions.
• Sequential Thinking — structured, step-by-step reasoning for harder multi-step tasks.
• Time — timezone and date conversions (boring, and surprisingly handy).

If you install nothing else, Filesystem + Git + Fetch covers the majority of day-to-day coding needs.

The integrations worth adding

Once the basics are in place, these pull their weight:

• GitHub — the official server lets an agent read code, manage issues and PRs, and run changes against a repo. If your work lives on GitHub, this is the highest-leverage add.
• Playwright (Microsoft) — give the agent a real browser to navigate, fill forms, and screenshot. The backbone of agentic end-to-end testing.
• A database server — let the agent inspect your schema and run read-only queries; enormous for "why is this query slow?" sessions. (Heads-up: the original reference Postgres and SQLite servers were archived in 2026, so reach for a maintained community or vendor server — e.g. your database provider's official MCP.)

How to vet a server before you trust it

An MCP server runs with real permissions on your machine or in your cloud. Treat installing one like adding a dependency — because that is exactly what it is.

1. Read what it can do. An MCP server declares its tools. Skim them. A "read-only docs" server that asks to run shell commands is a red flag.
2. Prefer least privilege. Point the Filesystem server at a project directory, not your home folder. Use read-only database credentials when you only need to inspect.
3. Check who maintains it. Official (modelcontextprotocol) and first-party vendor servers are safer bets than an anonymous one-off.
4. Pin the version. Don't auto-run the latest tag for something with filesystem or network access. Pin it, and read the changelog before bumping.

A minimal, sane config

Most clients use a JSON config that maps a server name to a command. A starter that is useful without being reckless:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/your/project"]
    },
    "git": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-git", "--repository", "/path/to/your/project"]
    }
  }
}

Note the explicit path: the Filesystem server only sees what you point it at.

The bottom line

The right number of MCP servers is "the fewest that unblock your actual work." Start with the official reference servers, add GitHub and a database server if you need them, and apply the same scrutiny you would to any dependency. Resist the urge to install a dozen at once — every server is surface area, and an agent with too many tools gets worse, not better, at choosing the right one.

Browse curated, free MCP servers — with install commands for each — in the AgentDepot directory →