The MCP servers worth installing first

There are hundreds of MCP servers, and every one you add is code that runs on your computer and tools your AI can call. Fewer is better. Here are the ones we'd start with, what each needs, and what to watch out for. Checked against the official docs in September 2026.

If you only install three

1. Context7: so the AI stops guessing at library APIs from stale memory.
2. Playwright: so the AI can check that a web page actually works.
3. GitHub or Filesystem: depending on whether your work lives on GitHub or on disk.

Code and docs

Context7

Pulls up-to-date documentation and code examples for libraries into the conversation. Works without an account; a free API key raises the rate limits.

Filesystem

Lets the AI read, search, edit and move files in the folders you name. Watch out: it can change files, so give it specific project folders, never your whole drive.

Git

Works with a local repository: status, diffs, history, and also staging, committing, resetting and switching branches. Needs uv. Watch out: it can write, not just read.

GitHub

GitHub's own server for repositories, issues and pull requests. It runs in Docker with a personal access token. Watch out: the token decides what the AI can do, so give it only the permissions you're comfortable with. Read-only is a good start.

Web

Playwright

Drives a real browser so the AI can open pages, click through them and take screenshots. Great for checking a UI change actually works. Watch out: only point it at sites you trust.

Fetch

Downloads a web page and converts it to readable text. Needs uv. Watch out: the maintainers warn it can reach local and internal network addresses, so keep that in mind on company networks.

Firecrawl

Scrapes and crawls websites into clean text. Needs a Firecrawl API key, so it relies on an outside service rather than running purely on your machine.

Notes and thinking

Notion

Search and edit pages and databases in your workspace. You create an internal integration in Notion and share only the pages you want the AI to see.

Memory

A simple knowledge graph the AI can store facts in, so it can carry context between conversations.

Sequential Thinking

Helps the AI break a hard problem into steps and revise them as it goes.

Servers to avoid

As of September 2026, several older "reference" servers published under @modelcontextprotocol on npm are marked as no longer supported. That includes the GitHub, Brave Search, Postgres and Slack packages. Guides written a year ago still tell you to install them. Use the current official server instead (for GitHub, the one above), and check a package's status on npm before you trust an old tutorial.

Keep your keys safe

Some servers need an API key or token. In Claude Code and Cursor, reference an environment variable in the config instead of pasting the key, so the file is safe to commit. Claude Desktop has no such option, so treat its config file as private. The builder does this for you.