The most useful MCP servers for everyday work are the reference servers from the Model Context Protocol project (files, Git, web fetch), GitHub's own server for repositories and issues, Microsoft's Playwright server for browser tasks, and Google's MCP Toolbox for databases. All four are open source on GitHub, and each installs with a single command or a short JSON block. This guide lists what each one does, how to install it, its license, and how to connect it to Claude Code without handing over more access than you need.

Who this is for and what you will have at the end

This is for developers, IT admins and technical founders who already use an AI assistant such as Claude Code and want it to touch real tools: a project folder, a Git history, a browser or a database. By the end you will have a short list of vetted repos, working install commands, a project-level .mcp.json you can share with a team, and a checklist for deciding whether to trust any other MCP server you come across.

If you want to build your own server instead of installing one, see Build Your First MCP Server in Python. If you are looking for Claude Skills rather than MCP servers, see Best Claude Skills GitHub Repos.

How MCP servers connect to your AI app

MCP is a protocol that lets an AI app call tools exposed by a separate program. There are two common ways a server is reached:

Transport How it runs Typical use
stdio The AI app starts the server as a local process and talks to it over standard input and output Local tools: files, Git, a local browser
HTTP (remote) The AI app connects to a URL Hosted servers, such as GitHub's

The Model Context Protocol README notes that servers default to stdio, so running one directly in a terminal looks like it hangs: it is waiting for a client to talk to it. That is normal.

Prerequisites

  • Claude Code installed. The commands below use its claude mcp command; the JSON blocks also work in most other MCP clients.
  • Node.js, for servers started with npx.
  • uv, for servers started with uvx (the Git and Fetch reference servers use it).
  • Docker, only if you choose the Docker route for the GitHub server.

How to add a server in Claude Code

Claude Code's MCP documentation gives two forms. Remote HTTP servers:

claude mcp add --transport http <name> <url> 

Local stdio servers:

claude mcp add [options] <name> -- <command> [args...] 

The -- matters for stdio servers. Per the documentation, it separates Claude's own options (--transport, --env, --scope) from the command that runs the server, and everything after it is passed to the server untouched.

Servers are stored by scope:

Scope Loads in Shared with team Stored in
Local (default) Current project only No ~/.claude.json
Project Current project only Yes, via version control .mcp.json in the project root
User All your projects No ~/.claude.json

Choose with --scope local, --scope project or --scope user. To see what is configured, run claude mcp list. Inside Claude Code, /mcp shows server status and handles sign-in for remote servers that use OAuth.

1. modelcontextprotocol/servers: the reference servers

What it does. This is the project's collection of reference servers. According to its README (checked 2026-11-19), the maintained ones are Everything (a test server), Fetch (web content fetching and conversion), Filesystem (file operations with configurable access controls), Git (read, search and manipulate repositories), Memory (a knowledge-graph memory), Sequential Thinking and Time.

Who it is for. Anyone starting with MCP. These are the most common first servers and the ones other guides assume you have seen.

Install commands from the README:

# Filesystem: only the folder you list is accessible npx -y @modelcontextprotocol/server-filesystem /path/to/allowed/files # Git uvx mcp-server-git # Fetch uvx mcp-server-fetch # Memory npx -y @modelcontextprotocol/server-memory 

Connecting the Filesystem server to Claude Code, limited to one folder:

claude mcp add --transport stdio files -- npx -y @modelcontextprotocol/server-filesystem /home/you/projects/contoso-site 

The folder argument is the access control. Give it the one project folder, not your home directory.

Windows note. The README says to wrap npx commands with cmd /c in client configs.

License. The README states Apache License 2.0 for new contributions, with existing code remaining under MIT. GitHub's license field for the repository reads NOASSERTION, because of that mix.

Repo facts (GitHub API, 2026-11-19): about 92,091 stars, last push 2026-11-18, not archived.

What moved out of this repo

Many servers people remember from older tutorials are no longer here. The README lists these as archived and moved to the servers-archived repository: AWS KB Retrieval, Brave Search, EverArt, GitHub, GitLab, Google Drive, Google Maps, PostgreSQL, Puppeteer, Redis, Sentry, Slack and SQLite. If a tutorial tells you to run @modelcontextprotocol/server-github or the PostgreSQL server, use the maintained replacements below instead. The README says Brave Search is replaced by the official @brave/brave-search-mcp-server and that Slack is now maintained by Zencoder.

2. github/github-mcp-server: repositories, issues and pull requests

What it does. GitHub's official MCP server lets an assistant work with repositories, issues and pull requests. Its default toolsets, per the README, are context, repos, issues, pull_requests and users.

Who it is for. Developers who want to ask an assistant to summarize an issue, review a pull request or search a repo without leaving the terminal.

Install options from the README:

  • Remote server hosted by GitHub, URL https://api.githubcopilot.com/mcp/, with "type": "http" in your host config. Authentication can be OAuth or a personal access token.
  • Docker, local:
docker run -i --rm \ -e GITHUB_PERSONAL_ACCESS_TOKEN=<your-token> \ ghcr.io/github/github-mcp-server 
  • A local binary built with go build -o github-mcp-server ./cmd/github-mcp-server, run as github-mcp-server stdio with the token set in the environment.

For Claude Code, the remote route follows the HTTP pattern above:

claude mcp add --transport http github https://api.githubcopilot.com/mcp/ 

Then run /mcp inside Claude Code to complete authentication. This command combines GitHub's URL with Claude Code's documented HTTP syntax; the exact authentication prompts depend on your account and Claude Code version, so check /mcp for status.

Limit what it can do. The README describes two controls. --toolsets repos,issues,pull_requests (or the GITHUB_TOOLSETS environment variable, which takes precedence) exposes only the toolsets you name. --read-only skips all write tools, and the README says it takes priority even over tools requested explicitly with --tools. For a first trial, enable read-only.

Token advice from the README. Minimum suggested scopes are repo, read:packages (for Docker image access) and read:org (for team access). It also recommends separate tokens per project, periodic rotation, and keeping tokens out of version control.

License. MIT. Repo facts (2026-11-19): about 34,576 stars, last push 2026-11-18, not archived.

3. microsoft/playwright-mcp: a real browser for the assistant

What it does. Playwright MCP lets an assistant drive a browser: open pages, read what is on them and interact. It is useful for checking a staging site, reproducing a front-end bug or filling a form during testing.

Who it is for. Developers and QA engineers testing web apps.

Install, from the README:

{ "mcpServers": { "playwright": { "command": "npx", "args": [ "@playwright/mcp@latest" ] } } } 

For Claude Code the README gives:

claude mcp add playwright npx @playwright/mcp@latest 

Security. The README says plainly that "Playwright MCP is not a security boundary." A browser that can open arbitrary pages can read content from those pages, and Claude Code's documentation warns that servers that fetch external content can expose you to prompt injection. Point it at sites you control, avoid logged-in sessions for sensitive systems, and do not combine it with a server that can write to production.

License. Apache-2.0. Repo facts (2026-11-19): about 39,868 stars, last push 2026-11-18, not archived.

4. googleapis/mcp-toolbox: databases

What it does. MCP Toolbox for Databases is an open source MCP server for databases. Per its README it offers prebuilt generic tools such as list_tables and execute_sql, and a framework for custom tools defined in a tools.yaml file. Listed supported databases include Google Cloud options (AlloyDB, BigQuery, Cloud SQL, Spanner, Firestore) and others such as PostgreSQL, MySQL, MariaDB, SQL Server, Oracle, MongoDB, Redis, Elasticsearch, CockroachDB, ClickHouse, Couchbase, Neo4j, Snowflake and Trino.

Who it is for. Analysts and developers who want an assistant to explore a schema and draft queries.

Install, from the README:

{ "mcpServers": { "toolbox-postgres": { "command": "npx", "args": [ "-y", "@toolbox-sdk/server", "--prebuilt=postgres", "--stdio" ] } } } 

Set the database connection environment variables described in the project's Prebuilt Tools Reference. For custom tools, the README shows npx @toolbox-sdk/server --config tools.yaml.

Rename. The repository was formerly googleapis/genai-toolbox (originally "Gen AI Toolbox for Databases") and is now googleapis/mcp-toolbox. Older tutorials and bookmarks may use the previous name. Update a local clone with git remote set-url origin https://github.com/googleapis/mcp-toolbox.git.

Safety. execute_sql can run whatever SQL the database user is allowed to run. Connect with a read-only database account, ideally against a replica or a copy, never a production primary with a write-capable user.

License. Apache-2.0. Repo facts (2026-11-19): about 16,905 stars, last push 2026-11-19, not archived.

Share a setup with your team

To give a whole project the same servers, add them with project scope so they are written to .mcp.json:

claude mcp add --transport http github --scope project https://api.githubcopilot.com/mcp/ 

The resulting file uses the format from Claude Code's documentation:

{ "mcpServers": { "shared-server": { "type": "http", "url": "https://example.com/mcp" } } } 

.mcp.json supports ${VAR} and ${VAR:-default} expansion in command, args, env, url and headers, so tokens can stay out of the committed file. Claude Code prompts for approval in interactive sessions before using project-scoped servers from .mcp.json.

Verify the setup

Run:

claude mcp list 

Claude Code's documentation says the list shows health statuses such as Connected, Needs authentication and Failed to connect. Inside a session, /mcp shows the same status and handles OAuth sign-in. Then ask the assistant for something harmless and observable, such as listing the files in the allowed folder or listing open issues in a public repo.

Troubleshooting

Symptom Likely cause Fix
Server runs in a terminal but nothing happens stdio servers wait for a client Run it through your MCP client instead
Status shows Needs authentication Remote server needs OAuth Run /mcp and sign in
Status shows Failed to connect Missing Node.js or uv, or a typo in the command Run the command by hand to see the error
npx fails on Windows Client cannot launch npx directly Wrap with cmd /c, as the README advises
CLI treats your server name as an environment variable --env swallowed the name Put another option, such as --transport stdio, between --env and the name
Old tutorial's server package not found Server was archived Use the maintained replacements above

Before you trust any other MCP server

  • Open its repository and check the license, the last commit date and whether it is archived.
  • Read what tools it exposes. A server that can write, delete or send is a larger risk than one that reads.
  • Prefer the vendor's own server (GitHub, Microsoft, Google) over a lookalike with a similar name.
  • Start with the narrowest scope and read-only credentials.
  • Remember that content fetched by a server can contain instructions aimed at the model.

Summary checklist

  • Start with Filesystem limited to one folder, plus Git.
  • Add GitHub's server in read-only mode before enabling writes.
  • Use Playwright MCP on sites you control only.
  • Use MCP Toolbox with a read-only database account.
  • Verify with claude mcp list and /mcp after each addition.

What we checked

  • Opened the README of modelcontextprotocol/servers on github.com on 2026-11-19 and copied the maintained-server list, the install commands and the archived-server list from it.
  • Queried the GitHub API for modelcontextprotocol/servers, github/github-mcp-server, microsoft/playwright-mcp and googleapis/mcp-toolbox on 2026-11-19: stars, last push date, license field and archived flag are as stated above. Star counts change daily.
  • Cross-checked license wording: the GitHub API reports NOASSERTION for modelcontextprotocol/servers, while its README states Apache-2.0 for new contributions and MIT for existing code.
  • Confirmed the repository rename from genai-toolbox to mcp-toolbox: the GitHub API returned the full name googleapis/mcp-toolbox when queried by the old name, and the README carries a rename notice.
  • Read Claude Code's MCP documentation (code.claude.com/docs/en/mcp) on 2026-11-19 for the claude mcp add syntax, scopes, .mcp.json format and the security warning. The page redirected from docs.claude.com.
  • Not tested hands-on: we did not run any of these servers or connect them to a live Claude Code session; the one command that combines GitHub's URL with Claude Code's HTTP syntax is assembled from the two documents, not run. The first 100,000 characters of the long Claude Code and GitHub README pages were read; the remainder was not.

Last tested: 2026-11-19

References

This guide was researched and drafted with AI assistance and checked against the vendor documentation listed in its References. Read the maintained version and related guides on EpiFive.