Best MCP servers for developers: what to install first
There are thousands of MCP servers. You need about four. The trick is not finding more; it is choosing the ones that remove a real chore, and installing them safely.
Pick by chore, not by hype#
A starter shortlist#
| Server | What it gives the model | Typical use |
|---|---|---|
| GitHub MCP server | Repositories, issues, pull requests | "Summarize the open bugs", "draft a PR description" |
| Playwright MCP | Drive a real browser | Test a UI flow, reproduce a front-end bug, check a page |
| Context7 | Up-to-date library documentation | Stop the model using an old API version |
| Filesystem (reference) | Controlled file read and write | Work on files inside an allowed directory |
| Git (reference) | Read, search and manipulate repos | History, diffs, blame |
| Fetch (reference) | Fetch and convert web pages | Read a docs page into context |
The last three come from the official modelcontextprotocol/servers repository, which also includes Everything (a test server), Memory, Sequential Thinking and Time. Its own README is clear that these are reference implementations and educational examples, not production-ready solutions, so read the code and limit access before you rely on them.
Install instructions change between versions and clients, so follow each project's README rather than a blog post.
Where to find more#
The official MCP Registry lists published servers, and the reference repository links to it. Browse it by the task you want to automate, then check the source.
How to vet a server in five minutes#
- Source. Is the repo maintained, with recent commits and real users? Who is the publisher?
- Tools. Read the list of tools and their descriptions. Do they match what the server claims, and nothing hidden?
- Permissions. What credentials does it need? Can you give it a read-only or narrowly scoped one?
- Transport. Local stdio runs as you. Remote HTTP needs authentication. See MCP security.
- Pin it. Use a specific version so an update does not change behavior under you.
Fewer is better#
Each server adds tool descriptions to the model's context. Twenty servers means hundreds of tools to choose from, more tokens spent, and more wrong picks. Keep the set small and focused, and disable the ones you do not use this week.
Build your own when it is simple#
If your chore is specific to your team (your deploy tool, your internal API), a small custom server is often better than a generic one. The MCP in Depth course builds one in TypeScript, and the first lesson, What is MCP?, takes about ten minutes.
New to the idea? Read MCP vs API: what is the difference?.
Frequently asked questions
What are the best MCP servers for developers?
Commonly useful ones are the GitHub server for repositories and issues, Playwright for browser automation, a documentation lookup server such as Context7, and the reference filesystem and git servers. Choose by the problem you want to stop doing by hand.
Where do I find MCP servers?
The official MCP Registry lists published servers, and the modelcontextprotocol/servers repository contains reference implementations. Always review a server before installing it.
Are the official reference MCP servers production-ready?
No. The repository describes them as reference implementations meant as educational examples, not production-ready solutions. Review and harden before relying on them.
How many MCP servers should I install?
Fewer is better. Every tool description takes space in the model's context and increases the chance it picks the wrong tool. Install the few you use every week.
Prefer plain text? Read this page as Markdown.