How to give Claude Code new tools with MCP

MCP servers are a kind of plug that lets Claude Code use other programs and services, such as GitHub or a browser. According to Anthropic's own documentation, you add a server with the command "claude mcp add," followed by a name and a web address or a command.

You can see your servers with "claude mcp list," get details with "claude mcp get" and remove a server with "claude mcp remove."

Three ways to connect

Claude Code supports three transport types, meaning ways of talking to the server: HTTP, which is recommended for cloud services, stdio for local programs on the computer, and WebSocket. The older SSE is marked as deprecated.

There are also three scopes, which determine who the server applies to. Local is the default and applies only to the current project without being shared. Project is shared with colleagues via the .mcp.json file. User applies across your own projects.

Some external servers require login via OAuth, a common login method. This is done with "claude mcp login" or via /mcp in a session. Alternatively, you can send a token in a header.

Which servers are worth starting with?

The Techsy guide, written by Mert Batur, recommends three servers first: Context7, which provides up-to-date documentation for code libraries, GitHub MCP for repositories, issues and pull requests, among other things, and Playwright, which can verify the work in a real browser. The guide also mentions Sequential Thinking, Filesystem, Supabase, Sentry, Exa, Memory, Serena and Sanity. The ranking is Techsy's own assessment.

Context7 and Playwright are free, according to the guide. Context7 is added with "claude mcp add context7 -- npx -y @upstash/context7-mcp," and the guide recommends user scope. Playwright is added with "claude mcp add playwright -- npx @playwright/mcp@latest," and here project scope is recommended for frontend projects.

Watch out for security

Anthropic warns that you should only add servers you trust. Servers that fetch external content can be vulnerable to prompt injection, where hidden instructions in the content try to control the AI.

Secrets in a shared .mcp.json can be specified as environment variables instead of in plain text. According to Techsy, servers from a project file are listed as "Pending approval" until each user has approved them.

Can you have many servers?

The author claims that the MCP Tool Search feature has been enabled by default since January 14, 2026, and cuts the AI's working memory usage by roughly 85-95 percent. As a result, the old rule of thumb of only 3-6 servers no longer applies, he argues. His own team runs 7 servers with 88 tools. The figure is the author's own claim and has not been independently verified.