The best Claude Code MCP servers, and how to choose
Straight up: the best Claude Code MCP servers are the ones that plug the agent into a tool or data source you already work with every day. An MCP server gives Claude Code a live connection to something outside itself, like your files, your repo, your database, or a browser. Without one, the agent can only reason about what you paste in. With the right ones, it can actually go and do the work.
So the choosing rule is simple. Install a server when you have a repeated task that needs that system, and skip it otherwise. If you do not touch a database, a Postgres server is dead weight. The list below is ranked by how often it pays off for a typical builder, not by how impressive it sounds. If you want the background on what MCP is first, we cover it in [what are MCP servers and why they matter](/blog/what-are-mcp-servers-and-why-they-matter).
The best Claude Code MCP servers, ranked by how often they earn their place
| MCP server | What it unlocks | Install it when |
|---|---|---|
| Filesystem | Read and write files in folders you choose | Almost always, for work outside the project dir |
| GitHub | Issues, pull requests, code, and reviews | You ship code or manage repos |
| Database (Postgres) | Query and inspect a real database | Your app has a database to work against |
| Browser (Playwright) | Drive a real browser and read live pages | You test, scrape, or automate the web |
| Fetch / web | Pull and read a specific URL on demand | The agent needs current, real web content |
| Slack | Read and post messages in channels | Your team or clients live in Slack |
| Google Drive | Read docs and sheets | Your source material sits in Drive |
| Memory | A persistent knowledge graph across sessions | You want recall that survives restarts |
Filesystem: the one almost everyone needs
Install the filesystem MCP server first. It lets Claude Code read and write files in folders you explicitly allow, beyond the single project it was launched in. That sounds small, but it is what turns the agent loose on real tasks: organizing a downloads folder, generating a batch of files, or working across two projects at once. You grant specific directories, so it stays scoped to what you intend.
GitHub: for anyone who ships code
Add the GitHub MCP server if you work with repositories. It connects Claude Code to issues, pull requests, code, and reviews, so the agent can open a PR, triage an issue, or read a repo it does not have checked out locally. For an agency, this is the difference between an assistant that suggests a fix and one that opens the pull request with the fix already in it.
Database and browser: where real apps live
Install a database server, such as the Postgres MCP server, when your work touches real data. It lets the agent inspect a schema, run read queries, and reason about your actual tables instead of guessing. Pair it with care: scope it to read-only or a safe database when you can, because a connected agent is a powerful one.
Add a browser server, such as the Playwright MCP server, when you need the agent to act on the live web. It can open pages, click, fill forms, and read what actually renders. That covers testing a flow, checking a deployed page, or pulling structured data from a site. It is the server that lets Claude Code see the web the way a user does, not just fetch raw HTML.
Fetch, Slack, Drive, and Memory: the situational picks
These four are worth it only when they map to your workflow. The fetch or web server pulls a specific URL on demand, which is useful when the agent needs current content from a page you name. The Slack server reads and posts messages, which pays off when your team or clients coordinate there. The Google Drive server reads docs and sheets, which helps when your source material lives in Drive. The memory server keeps a persistent knowledge graph across sessions, so the agent recalls facts after a restart.
How to install an MCP server in Claude Code
Installing an MCP server is a config step, not a coding project. You register the server with Claude Code, then restart so it loads. There are two common paths, and both take a few minutes.
- Pick the server and find its command. Most servers tell you the exact run command and any keys or tokens they need.
- Add it to your MCP config. Use the Claude Code command to add a server, or edit the config file the desktop app reads.
- Provide credentials as needed. GitHub, Slack, and database servers want a token or connection string, kept in your config or environment.
- Restart Claude Code so it picks up the new server, then ask the agent to use it on a small task to confirm it connected.
How many MCP servers should you run?
Run as few as cover your real work. More servers mean more context, more credentials to manage, and more surface area for the agent to misfire on. Four well-chosen servers you call often beat a dozen you installed once and forgot. This is the same principle behind picking [the Claude Code skills to install first](/blog/claude-code-skills-to-install-first): match the tool to the work in front of you, then add more only when a new repeated task demands it.
The agency angle: which MCP servers earn their keep on client work
For client work, the servers that pay are the ones that compress delivery time. GitHub turns review notes into shipped pull requests. A database server lets you answer data questions without a round trip to the client's engineer. A browser server lets you verify a live deployment before you tell the client it is done. Each one removes a handoff, and removed handoffs are where agency margin comes from.
Price that leverage deliberately. When MCP servers let you deliver in hours what used to take days, that is value to bill for, not a discount to pass on. We break down how to think about it in [how to price a Claude Code agency project](/blog/how-to-price-a-claude-code-agency-project).
Frequently asked questions
What is the best MCP server to install first in Claude Code?
The filesystem MCP server. It lets Claude Code read and write files in folders you allow, which is the foundation for almost every real task. After that, add GitHub, a database server, and a browser server based on the work you do.
What does an MCP server do in Claude Code?
An MCP server gives Claude Code a live connection to a tool or data source outside itself, like your files, a repo, a database, or a browser. It lets the agent act on real systems instead of only reasoning about text you paste in.
How do I install an MCP server in Claude Code?
Register the server with Claude Code using its add-server command or by editing the config file the desktop app reads, provide any tokens it needs, then restart Claude Code. Test it on a small task to confirm the connection.
How many MCP servers should I run at once?
As few as cover your real work, often around four. Each server adds context and credentials to manage, so a small set you use often beats a long list you installed and forgot.
Are MCP servers safe to use?
They are as safe as the access you grant. A connected server can act on the real system, so scope tokens to the minimum the task needs, prefer read-only access where possible, and never connect write access to something you cannot afford to have changed by accident.
Which MCP servers are best for agency and client work?
GitHub, a database server, and a browser server tend to earn their keep, because they turn review notes into shipped pull requests, answer data questions without a client round trip, and verify live deployments. Each removes a handoff, which is where agency margin comes from.
Last reviewed by Duncan Rogoff on June 22, 2026
