Best MCP Servers for Claude in 2026: 8 Worth Installing (and How to Add Them)
Quick answer: Start with a small set: the GitHub MCP server for repos and pull requests, Playwright for browser testing, Context7 for up-to-date library docs, plus the reference Filesystem, Fetch and Memory servers.
Add one server at a time, give each the least access it needs, and read its license before you rely on it.
MCP (Model Context Protocol) is the open standard that lets Claude talk to outside tools and data. An MCP server is the piece that exposes a service, such as GitHub or a browser, to Claude in a form it can call.
There are thousands of servers, and most lists are just long. This guide keeps it short. Every server below was checked against its own repository on November 18, 2026, including the license and the install command it documents.
How we picked these servers
We kept servers that meet three tests:
- A maintained repository. Each repo had a commit within the last few weeks when we checked.
- A clear license. We name the license for each one, because it matters if you ship software to customers.
- A documented install path for Claude. Either a
claude mcp addcommand in the repo or a standard config block.
We did not rank by popularity, and we do not quote star counts, because they change daily and we could not verify them in one consistent way.
The 8 best MCP servers for Claude
| Server | What it gives Claude | License (checked Nov 18, 2026) | Best for |
|---|---|---|---|
| GitHub MCP server | Repos, issues, pull requests, code search | MIT | Any developer on GitHub |
| Playwright MCP | Real browser control and page snapshots | Apache-2.0 | Testing and web automation |
| Context7 | Current docs for libraries and frameworks | MIT | Fewer outdated code suggestions |
| Filesystem (reference) | Scoped file access with allowed directories | MIT to Apache-2.0 transition* | Giving Claude a defined workspace |
| Fetch (reference) | Web page content converted for LLM use | MIT to Apache-2.0 transition* | Reading docs and articles |
| Memory (reference) | Knowledge-graph memory across sessions | MIT to Apache-2.0 transition* | Persistent project facts |
| Sequential Thinking (reference) | Structured, step-by-step reasoning | MIT to Apache-2.0 transition* | Hard multi-step problems |
| Notion MCP server | Read and write Notion pages and databases | MIT | Teams that live in Notion |
*The official modelcontextprotocol/servers repository says the project is moving from the MIT License to Apache-2.0. New contributions are Apache-2.0, so check the license file for the exact server version you install.
1. GitHub MCP server
GitHub's own server lets Claude read repositories, search code, and work with issues and pull requests. It is the most useful single server for most US dev teams, because it turns "review this PR" into a task Claude can complete end to end.
The repo documents a Docker-based setup for Claude Code:
claude mcp add github -e GITHUB_PERSONAL_ACCESS_TOKEN=YOUR_GITHUB_PAT -- docker run -i --rm -e GITHUB_PERSONAL_ACCESS_TOKEN ghcr.io/github/github-mcp-server
Use a fine-grained personal access token limited to the repositories you want Claude to touch. A broad token is the most common mistake we see.
2. Playwright MCP
Microsoft's Playwright server gives Claude a real browser. It can open pages, click, fill forms and read the page structure, which makes it a strong fit for testing a signup flow or checking a deploy. The repo documents this one-liner:
claude mcp add playwright npx @playwright/mcp@latest
Keep it pointed at your own staging sites and test accounts. Do not hand it a logged-in browser profile that holds personal or banking sessions.
3. Context7
Context7 solves a specific problem: models suggest code for older versions of a library. The server exposes two tools. resolve-library-id finds the right library, and query-docs pulls current documentation for it. Its README describes a guided setup:
npx ctx7 setup
It is a good first add if you work with fast-moving frameworks. It only reads public documentation, so the risk is low.
4. Filesystem (reference server)
The Filesystem server gives Claude file operations with configurable access controls. The point is the allow-list: you name the directories Claude may use, and it stays inside them.
Anthropic's MCP team labels the servers in this repository as reference implementations meant to teach and demonstrate. They are not described as production-ready, so treat them as a starting point and review the source if you plan to depend on one.
5. Fetch (reference server)
Fetch retrieves a web page and converts it into content that is efficient for a model to read. It is handy for pulling in documentation or an article without pasting it by hand.
Remember that fetched pages are untrusted input. A page can contain text that tries to steer the model. Read our note on this under the security section below.
6. Memory (reference server)
Memory stores facts in a knowledge graph so a project can remember decisions between sessions. It is simple and local. If you want to compare heavier options, our roundup of best GitHub repos for AI agent memory covers them, and the agent memory MCP skill shows one packaged approach.
7. Sequential Thinking (reference server)
This server structures reasoning as a sequence of thoughts that can be revised. It helps on planning-heavy work like architecture decisions. It adds tokens and time, so skip it for simple edits.
8. Notion MCP server
Notion's official server is MIT-licensed and lets Claude work with pages and databases. It suits US startups that keep specs and meeting notes in Notion. Grant it access only to the workspace pages Claude needs, not the whole workspace.
Servers we left off, and why
- Sentry MCP. It is useful, but its repository uses the Functional Source License, which is source-available rather than a standard open-source license. Read the terms before you build on it.
- Archived reference servers. The official repository lists several former reference servers as archived, including GitHub, PostgreSQL, Puppeteer and Slack. Some have official replacements maintained elsewhere. Prefer the maintained version.
How to add an MCP server to Claude Code
The general pattern is the same for every server:
claude mcp add <name> -- <command to start the server>
claude mcp list
Then restart your session if needed and ask Claude to use the new tools. Servers can be configured for just you or shared with your project, so decide which fits before you commit a config file to a repo.
Security habits for US teams
An MCP server runs with whatever access you give it, so treat each one like a new contractor with a key card.
- Least privilege. Use scoped tokens, read-only access where possible, and separate test accounts.
- Read the source or the license. Know who maintains it and how it handles your data.
- Treat tool output as untrusted. Web pages, issues and documents can contain instructions aimed at the model. Do not let a server that reads outside content also hold powerful write access without review.
- Check your compliance needs. If your company is subject to SOC 2 or customer data rules, run a new server past whoever owns that review before connecting it to production data.
For a deeper review process, see the MCP server security skill, and browse the full AI Agents and MCP category.
MCP servers vs. skills
People often ask whether they need servers or skills. The short version: a server gives Claude access to a service, and a skill teaches Claude a method. Most real workflows use both. Our guide to Claude skills vs MCP explains when each one wins, and what are Claude skills covers the basics. If you want to build your own server, the MCP builder skill is a good starting point.
FAQ
What is the best MCP server for Claude? For most developers, the GitHub MCP server. If you test web apps, add Playwright. If you work with changing libraries, add Context7.
Are MCP servers free? The servers listed here are open source or source-available and free to run. The services they connect to, such as GitHub or Notion, have their own pricing.
Is it safe to install any MCP server? No. Install only servers you understand, give them scoped credentials, and keep write access narrow.
Do I need MCP if I use Claude skills? Not always. Skills cover repeatable methods. You need MCP when Claude has to reach a live service or tool.
How many MCP servers should I run at once? Start with two or three. Each server adds tools Claude has to consider, and more servers mean more to secure.
Sources: modelcontextprotocol/servers, github/github-mcp-server, microsoft/playwright-mcp, upstash/context7, makenotion/notion-mcp-server, getsentry/sentry-mcp.