AI Coding Guide

Five MCP servers worth installing first (Claude Code, Codex, Gemini CLI)

General Published:

Five MCP servers that pay off immediately in all three coding agents, what each is for, the install command, and why installing too many costs you context on every request.

Verified on Sep 7, 2026 These tools change quickly. Please also check the latest official documentation.
#MCP#Playwright#Context7#GitHub#recommendations
Contents
  1. How these were chosen
  2. 1. Playwright: drive a browser and look at the result
  3. 2. Context7: read a library's current documentation
  4. 3. GitHub: read and write issues and pull requests
  5. 4. Filesystem: expose exactly the directories you choose
  6. 5. Sequential Thinking: break a hard problem into steps
  7. Why a long list costs you
  8. Per-tool setup
  9. Summary

Hundreds of MCP servers exist, but only a handful make a difference you can feel in day-to-day coding work. This article narrows it to five that work in Claude Code, Codex and Gemini CLI alike and are useful from the moment you install them.

KEY POINT

What you will learn

  • Five MCP servers worth installing first, and what each is for
  • Install commands that work across all three tools
  • Why a long server list is a cost, not a feature

How these were chosen

  • They supply what the model cannot get on its own — what a page actually renders, what the current version of a library documents
  • They are official or widely used, so maintenance continues
  • They are read-leaning, so mistakes are cheap, and write tools can be narrowed with approval rules

1. Playwright: drive a browser and look at the result

The highest-impact server for web work. The agent opens your dev server, takes screenshots, clicks buttons and reads what's on screen — which removes the "implemented it but never looked at it" state.

# Claude Code
claude mcp add playwright -- npx @playwright/mcp@latest
# Codex (~/.codex/config.toml)
[mcp_servers.playwright]
command = "npx"
args = ["@playwright/mcp@latest"]
// Gemini CLI (.gemini/settings.json)
{ "mcpServers": { "playwright": { "command": "npx", "args": ["@playwright/mcp@latest"] } } }

A prompt that uses it: "Start the dev server, open /login, submit the empty form and tell me what the error state looks like."

The browser is real

Playwright drives an actual browser. Keep it pointed at development URLs, and don't hand it production URLs or a logged-in session.

2. Context7: read a library's current documentation

For a library version newer than the model's training data, or one whose API changes often, this fetches the relevant documentation and puts it in front of the model. It is the direct answer to "it wrote the old API".

claude mcp add --scope user context7 -- npx -y @upstash/context7-mcp

A prompt that uses it: "Check the current Prisma docs with Context7, then add the composite unique constraint."

It is read-only, so auto-approving it is safe.

3. GitHub: read and write issues and pull requests

For reading an issue before planning a fix, or reading a PR's diff and review comments. If gh is available it covers most of this; the MCP server earns its place where shell commands aren't allowed, or where a Codex sandbox restricts the network.

claude mcp add github -e GITHUB_PERSONAL_ACCESS_TOKEN=${GITHUB_TOKEN} -- npx -y @modelcontextprotocol/server-github

Pass the token through an environment variable rather than writing it into a config file. A good split is to auto-approve the read tools and leave issue and PR creation on ask.

用語解説

The official GitHub MCP server: GitHub itself offers a remote (HTTP) MCP server. Instead of running an npm package locally, you point at a URL and authenticate in the browser. Check each tool's documentation for its support.

4. Filesystem: expose exactly the directories you choose

A coding agent normally only touches files in its working directory. When you want it to read another repository's implementation or a shared documentation folder, this server exposes only the directories you name.

claude mcp add docs-fs -- npx -y @modelcontextprotocol/server-filesystem /Users/me/shared-docs

Nothing outside the directories you pass as arguments is reachable, which makes it a good fit for deliberately narrowing what the model can read.

5. Sequential Thinking: break a hard problem into steps

This one supplies a method rather than data: it walks the model through decomposing a problem for design decisions and difficult debugging. With reasoning built into current models the benefit varies by model and task, but it is lightweight and has no side effects, so it is cheap to try.

claude mcp add sequential-thinking -- npx -y @modelcontextprotocol/server-sequential-thinking

Why a long list costs you

Every connected server's tool definitions consume context on every request, used or not. Keep ten or more servers connected and you are sending thousands to tens of thousands of tokens before you type anything.

  • Put only what a project uses daily in project scope, where the team shares it
  • Keep occasional servers in local or user scope and remove them when the need passes
  • In Claude Code, /context shows how much space tool definitions occupy — see Managing context in Claude Code

Per-tool setup

For install commands and configuration files in detail:

Summary

  • Start with Playwright (see the screen) or Context7 (current docs); both pay off immediately
  • Auto-approve GitHub's read tools, keep writes on ask, and pass the token through the environment
  • Filesystem is for narrowing what gets read; Sequential Thinking is free to try
  • Unused servers only cost context, so start with two or three and grow from need

FAQ

Should I install as many MCP servers as possible?
No. Every connected server's tool definitions consume context whether or not they get used. Start with the two or three you actually use and add more when a need appears.
Can one MCP server work in all three tools?
Yes. The start command and arguments are the same; only the configuration format differs (JSON or TOML).
GitHub MCP server or the gh command?
If gh is installed, it covers most of what you need. The MCP server earns its place where you don't want to allow shell commands, or where you want structured results.

Primary sources

This article was drafted by AI from official documentation and reviewed by the site operator before publishing. Found a mistake? Let us know via the contact page.

Related articles