The best MCP servers for Claude Code
MCP (Model Context Protocol) servers give Claude Code new tools: your issue tracker, a real browser, error logs, up-to-date docs. The right three or four make it noticeably more useful; twenty make it slower and more confused. Here's a short list of well-known, actively used servers, how to add them, and how to run your own as a remote MCP server.
The list
| Server | From | What it's good for | Runs |
|---|---|---|---|
| GitHub | GitHub (official) | Issues, pull requests, code search, reviews and CI status without leaving Claude Code. | local or remote (hosted by GitHub) |
| Playwright | Microsoft (official) | Drives a real browser through accessibility snapshots: click through your app, fill forms, check what a page actually shows. | local |
| Context7 | Upstash | Pulls current, version-specific library documentation into context, so the model stops guessing at APIs that changed. | local or remote |
| Sentry | Sentry (official) | Reads errors, stack traces and issue details, so Claude Code can go from a production error to a fix. | remote |
| Atlassian (Jira & Confluence) | Atlassian (official) | Reads and updates Jira tickets and Confluence pages — pick up a ticket, write the code, update the status. | remote |
| Linear | Linear (official) | Issues and projects in Linear, for teams that track work there instead of Jira. | remote |
| Notion | Notion (official) | Searches and edits Notion pages and databases — specs, docs, meeting notes. | remote |
| Figma | Figma (official) | Gives the model design context from Figma files — layout, components, variables — to implement a design closely. | local (Figma desktop app) or remote |
| Slack | Slack and community servers | Reads channels and threads and posts messages — useful for summaries and notifications. | varies by server |
| Databases (PostgreSQL and others) | various, e.g. Supabase | Lets the model inspect schemas and run queries. Use a read-only database user unless you really mean otherwise. | local or remote |
| Stripe | Stripe (official) | Works with Stripe objects and documentation — handy when building payments. | local or remote |
| Cloudflare | Cloudflare (official) | A set of servers for Cloudflare's platform: docs, Workers, logs and more. | remote |
| Filesystem, Fetch, Memory, Git | MCP reference servers | The protocol's reference servers. Claude Code already reads files and fetches pages itself, so these matter more for other clients — Memory (a persistent knowledge graph) is the one most often added. | local |
Our picks by job
For almost everyone: GitHub
If your code lives on GitHub, this is the first one to add. Claude Code can read the issue it's fixing, open the pull request and check CI — the loop you'd otherwise do by hand in a browser tab.
For frontend work: Playwright
Without a browser, the model can only assume what your page looks like. With Playwright it can load the page, click through a flow and see what broke. It's also a good way to check a site after deploying it.
For fast-moving libraries: Context7
Models know libraries as of their training data. A docs server fetches the current documentation for the version you use, which cuts down on calls to functions that were renamed or removed.
For production bugs: Sentry
Point Claude Code at an issue and it can read the stack trace and context directly, then go and find the code.
For teams: your tracker and wiki
Jira and Confluence (Atlassian), Linear or Notion — whichever your team uses. The value is in closing the loop: read the ticket, do the work, update the ticket.
For data: a database server
Letting the model look at the real schema beats pasting it into the chat. Connect with a read-only user, and point it at a copy rather than production when you can.
How to add an MCP server to Claude Code
Claude Code manages servers with the claude mcp command. The general shape:
# a local (stdio) server: Claude Code starts the command itself
claude mcp add my-server -- <command> [args...]
# a remote server over HTTP
claude mcp add --transport http my-server https://example.com/mcp
# see what's configured
claude mcp list- Remote servers that need a token can take it as a header (
--header "Authorization: Bearer …"); many official remote servers use an OAuth login instead, which you complete from the/mcpmenu inside Claude Code. - Servers can be configured for just you or shared with your team through a project file (
.mcp.json) — see the--scopeoption. - The exact command for each server is in its README. Copy it from there rather than from a blog post.
Local vs remote MCP servers
| Local (stdio) | Remote (streamable HTTP) | |
|---|---|---|
| Where it runs | On your machine, started by Claude Code | On a server, reached over HTTPS |
| Who can use it | One client on one machine | Any client, any teammate, agents running elsewhere |
| Setup | Install the runtime and the package locally | Just a URL and a token or login |
| Good for | Things that need your local files or desktop apps | Shared tools, SaaS integrations, anything you'd otherwise install on every laptop |
Most SaaS vendors now ship remote MCP servers — you connect to their URL and log in, and there's nothing to install. For your own tools, a remote server means you write it once and the whole team uses it.
Keep it safe and fast
- Fewer is better. Every server's tool descriptions take up context. Enable the ones you use, remove the rest.
- Least privilege. Scoped tokens, read-only database users, one repository rather than all of them.
- Mind prompt injection. Tool results — an issue body, a web page, a Slack message — are text the model reads. Treat servers that pull in untrusted content with the same care as browsing.
- Install from the source. Prefer the vendor's official server or a well-known project over a random fork.
Host your own remote MCP server
The best MCP server for your project is often the one you write: your internal API, your database, your deploy script, wrapped as a few tools. Building one takes a few lines of Python — here's how to build an MCP server. To make it a remote MCP server everyone can use, it needs a long-running process with HTTPS and a token. Paste this to Claude Code from your server's folder:
It creates a server on The Vibe Hosting, starts your MCP server over streamable HTTP and gives you a URL to add with
claude mcp add --transport http. A small MCP server running 24/7 fits inside the free monthly credit of a
verified account. Details and prices: MCP server hosting ·
host a remote MCP server.
FAQ
What are the best MCP servers for Claude Code?
For most developers: GitHub for issues and pull requests, Playwright for browser testing, a docs server such as Context7, and Sentry for production errors. Add your team's tracker (Jira, Linear or Notion) and a read-only database server as needed.
How do I add an MCP server to Claude Code?
Use the claude mcp add command — with the server's start command for a local stdio server, or with --transport http and a URL for a remote one. claude mcp list shows what's configured.
What is the difference between local and remote MCP servers?
A local server runs on your machine and is started by the client over stdio. A remote MCP server runs on a server and is reached over HTTPS, so many clients and people can share it.
How many MCP servers should I use?
As few as do the job. Each server's tools take up context, so three or four focused servers usually beat a long list.
Can I host my own remote MCP server for free?
Yes. The Vibe Hosting runs it on a real Linux server with HTTPS; a small server fits inside the free monthly credit, and your coding agent deploys it with one prompt.
Try it: paste this to Claude Code, Cursor or Codex.