Best MCP Servers for Claude Code (2026)

There are over 17,000 MCP servers indexed across various registries right now. That number was around 100 eighteen months ago.

Most of those 17,000 aren't worth your time. Some are proof-of-concept demos that never got maintained. Some bind to 0.0.0.0 and expose your entire system to the internet (security researchers flagged hundreds of these in mid-2025). And a good chunk are near-duplicates of each other — you'll find a dozen different "file system" servers with varying levels of polish.

We spent the last month testing over 40 MCP servers across development, content, and operations workflows. These are the 12 that actually changed how we work — and the honest tradeoffs for each.

What We Mean by "Best"

Not just "most popular." Not just "most GitHub stars." We evaluated based on four things:

• Does it save real time? Not theoretically. Did it measurably reduce steps in an actual workflow?
• Is it maintained? When was the last commit? Does it work with current Claude Code versions?
• Is it secure? Does it follow sensible defaults — scoped permissions, no unnecessary network exposure?
• Is it stable? Did it crash, timeout, or produce inconsistent results during testing?

A server that checks all four is rare. Most check two or three. We'll be specific about which ones and where.

Quick Reference: Our Picks

Server	Best For	Setup Difficulty	Maintained
GitHub	PR reviews, issue management, CI/CD	Easy	Yes (official)
Context7	Up-to-date library docs	Easy	Yes
Sequential Thinking	Complex multi-step reasoning	Easy	Yes (official)
Playwright	Browser testing, web automation	Easy	Yes (official)
PostgreSQL	Database queries, schema exploration	Medium	Yes
Brave Search	Web research inside Claude	Easy	Yes
Sentry	Error triage and debugging	Medium	Yes (official)
File System	Local file operations beyond defaults	Easy	Yes (official)
Replicate	AI image/video/audio generation	Easy	Yes (official)
Figma	Design-to-code translation	Medium	Yes (official)
Notion	Knowledge base, project docs	Medium	Yes
Linear	Issue tracking, sprint management	Easy	Yes (official)

Now the details.

Tier 1: Install These First

These four servers improve nearly every Claude Code session, regardless of what you're building.

1. GitHub MCP Server

What it does: Connects Claude to GitHub's API — issues, pull requests, commits, Actions workflows, code search. You interact with repos through conversation instead of switching to the browser.

Why it made the list: This is the most universally useful MCP server we tested. The workflow improvement is immediate and concrete. Instead of "let me check GitHub" → copy the issue text → paste into Claude → ask for help, you say "look at issue #247 and suggest a fix." Claude reads the issue, checks the referenced code, and proposes changes. One step instead of five.

PR reviews became faster too. Claude reads the diff, cross-references with the codebase, and flags potential issues with specific line references. We caught a race condition in a review that the human reviewer had missed — not because the human was careless, but because Claude could hold the entire diff and the relevant source files in context simultaneously.

Setup:

claude mcp add github --scope user

You'll need a GitHub personal access token with appropriate scopes. Keep permissions minimal — read-only for repos you don't want Claude modifying.

Honest tradeoffs: The token-based auth means you're giving Claude access to your GitHub data. Scope it tightly. Don't use a token with org-admin permissions for a coding assistant. Also, large monorepos can hit rate limits during heavy use.

2. Context7

What it does: Fetches current documentation for libraries and frameworks in real time. When Claude needs to reference React, Next.js, Tailwind, or any of the thousands of supported libraries, it pulls the latest docs rather than relying on training data.

Why it made the list: This solved a problem we hit constantly: Claude confidently using deprecated API patterns. Before Context7, about 1 in 5 code suggestions referenced outdated methods — not wrong enough to throw errors immediately, but wrong enough to cause subtle bugs later. With Context7, Claude checks the current docs first.

The difference is most visible with fast-moving libraries. Anything that ships major updates quarterly (Next.js, Prisma, tRPC) benefits significantly. For stable APIs that haven't changed in years, the impact is smaller.

Setup:

claude mcp add context7 -- npx -y @upstash/context7-mcp@latest

No API key required. It works out of the box.

Honest tradeoffs: Adds latency to responses that trigger doc lookups — usually 1-3 seconds. If you're working with a library that has poor or sparse documentation, Context7 can't fix that. It's fetching what exists.

3. Sequential Thinking

What it does: Gives Claude a structured reasoning process for complex problems. Instead of jumping straight to an answer, Claude breaks the problem into steps, reasons through each one, revises its approach when something doesn't hold, and tracks its thinking chain.

Why it made the list: The output quality improvement on multi-step problems is noticeable. We tested it on architecture decisions, debugging sessions, and refactoring plans. Without Sequential Thinking, Claude tends to propose the first reasonable solution. With it, Claude considers alternatives, identifies edge cases it would have missed, and often arrives at a better answer.

One example: we asked Claude to refactor a payment processing module with several edge cases around currency conversion and retry logic. Without Sequential Thinking, the proposed refactor missed a retry scenario that would have double-charged customers. With it, Claude identified the scenario during its reasoning chain and handled it in the design.

Setup:

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

Honest tradeoffs: It makes responses slower — sometimes significantly. For straightforward tasks (rename a variable, fix a typo, write a simple function), the extra reasoning overhead doesn't add value and just costs time and tokens. Use it for problems where you'd want a senior engineer to think carefully before answering, not for everything.

4. Playwright

What it does: Gives Claude the ability to interact with web applications through browser automation. Navigate pages, fill forms, click buttons, take screenshots, extract data — all based on accessibility trees rather than pixel analysis.

Why it made the list: Two words: test automation. Claude can write and run browser tests, then iterate based on actual results. Before Playwright MCP, writing E2E tests with Claude was a two-step process — Claude writes the test, you run it, paste the failure, Claude fixes it, repeat. With Playwright MCP, Claude writes the test, runs it, sees the failure, and fixes it in one loop.

We also used it for scraping documentation pages and extracting structured data. Claude navigates the site, identifies the relevant content using accessibility selectors (not brittle CSS paths), and pulls exactly what you need.

Setup:

claude mcp add playwright -- npx -y @anthropic/mcp-playwright

Honest tradeoffs: Browser automation is resource-heavy. Running Playwright sessions eats RAM and CPU. On a machine with less than 16GB of RAM, concurrent browser sessions can slow everything down. Also, some sites actively block automated browsers, so scraping isn't always reliable.

Tier 2: Add Based on Your Workflow

These servers are excellent but situational. Install them when they match what you're building.

5. PostgreSQL

What it does: Connects Claude to your PostgreSQL databases. Run queries, explore schemas, analyze data — in natural language. Claude translates your questions into SQL, executes them, and explains the results.

Why it made the list: For teams that work with Postgres daily, this eliminates the context switch between your terminal and a database client. Ask Claude "which users signed up in the last 7 days and haven't completed onboarding?" and it writes and runs the query. The SQL it generates is surprisingly good — proper JOINs, appropriate indexes referenced, correct date handling.

Setup:

claude mcp add postgres -- npx -y @modelcontextprotocol/server-postgres postgres://user:password@localhost:5432/mydb

Honest tradeoffs: You're giving Claude direct database access. Use a read-only connection string for analysis work. We wouldn't point this at a production database with write permissions — one misunderstood instruction and Claude could modify or delete data. Test environments and analytics replicas are the safe bet.

6. Brave Search

What it does: Web search from within Claude Code. When Claude needs current information — latest package versions, recent security advisories, current pricing, live documentation — it searches the web and incorporates the results.

Why it made the list: Claude's training data has a cutoff. When you're debugging an error message from a library that shipped last week, Claude without web search can only guess. With Brave Search, it finds the relevant GitHub issue, Stack Overflow thread, or changelog entry and gives you a useful answer.

We found it most valuable during debugging. A cryptic error message that Claude couldn't explain from memory was usually solved within seconds once it could search for the exact error string.

Setup:

claude mcp add brave-search -- npx -y @anthropic/mcp-brave-search

Requires a Brave Search API key (free tier available at brave.com/search/api).

Honest tradeoffs: Search results vary in quality. Claude sometimes pulls in SEO-optimized content that sounds confident but is wrong. It's a good research tool, not an oracle. Verify important findings.

7. Sentry

What it does: Connects Claude to your Sentry error tracking. Fetch recent errors, analyze stack traces, check error trends, and investigate production issues without leaving your terminal.

Why it made the list: Error triage is one of those tasks that eats hours before you even start fixing anything. Opening Sentry, reading the stack trace, finding the relevant code, understanding the context — Claude can now do that entire flow in one step. "What's the most frequent error this week and where is it coming from?" gets you straight to the fix.

Setup:

claude mcp add sentry -- npx -y @sentry/mcp-server

Requires your Sentry auth token and organization slug.

Honest tradeoffs: Only useful if you're already using Sentry. The server handles standard error tracking well but doesn't cover all Sentry features — performance monitoring and session replay aren't fully exposed through the MCP interface yet.

8. File System

What it does: Extends Claude Code's file operations with configurable access controls. Read, write, search, and manage files across specified directories with explicit permission boundaries.

Why it made the list: Claude Code already handles files in your project directory. The File System MCP server extends that to other directories you specify — shared config folders, documentation repos, asset directories. The access control model lets you whitelist specific paths, so Claude can read your design tokens folder without having access to your entire home directory.

Setup:

claude mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem /path/to/allowed/directory

Honest tradeoffs: The default Claude Code file handling covers most cases. This server adds value when you need cross-directory access or explicit permission boundaries. If your workflow is "work in one project at a time," you might not need it.

Tier 3: Specialist Picks

These servers are excellent for specific roles or workflows.

9. Replicate

What it does: Gives Claude access to Replicate's API — thousands of AI models for image generation, video creation, audio processing, and more. Generate images, upscale photos, create videos, run inference on specialized models — all from conversation.

Why it made the list: This is the server nobody else is talking about. If your workflow involves any visual content creation, Replicate MCP turns Claude from a text-only assistant into a multimodal creative tool. Need a product shot? Claude generates it. Need a social media graphic? Claude creates it. Need to upscale a low-res image? Claude runs it through an upscaler model.

We used it to generate all the social media graphics for a product launch. The workflow was: describe what we need → Claude generates it via Replicate → review → iterate. No separate design tool, no switching apps.

Setup:

claude mcp add replicate -- npx -y mcp-replicate

Requires a Replicate API token. Costs depend on which models you run — most image generation models cost $0.01-0.05 per run.

Honest tradeoffs: Replicate usage costs money per API call. For heavy image generation workflows, costs add up. Also, AI-generated images still need human review — especially for brand-sensitive content.

10. Figma

What it does: Connects Claude to Figma files. Read design specifications, extract component properties, translate designs into code. The design-to-code pipeline, streamlined.

Why it made the list: If you implement designs from Figma, this is a significant workflow improvement. Claude reads the actual design file — spacing values, color tokens, component hierarchy — rather than you describing it manually or screenshotting. The resulting code is more accurate because it works from the source data, not your interpretation of it.

Setup: Uses OAuth for authentication. Follow the setup at Figma's MCP documentation.

Honest tradeoffs: Works well for component-level design translation. Less useful for complex multi-page layouts where the design-to-code mapping requires subjective decisions about responsiveness and interaction patterns. Claude gives you a starting point, not a pixel-perfect implementation.

11. Notion

What it does: Connects Claude to your Notion workspace. Search pages, read content, update documents, manage databases. Your knowledge base becomes accessible from Claude Code.

Why it made the list: If your team stores documentation, specs, or project notes in Notion, this saves real context-switching time. "Pull the API spec from our Notion docs and implement the endpoint" is a single instruction instead of: open Notion, find the page, copy the relevant section, paste into Claude.

Setup:

claude mcp add notion -- npx -y @anthropic/mcp-notion

Requires a Notion integration token with access to your relevant workspaces.

Honest tradeoffs: Notion's API has rate limits that can slow things down during heavy use. Also, giving Claude access to your entire Notion workspace might expose sensitive internal documents — scope the integration to specific databases or pages where possible.

12. Linear

What it does: Connects Claude to Linear for issue tracking and project management. Read issues, create tasks, update statuses, comment on threads — all from your terminal.

Why it made the list: For dev teams using Linear, the "read issue → understand context → start working" loop gets compressed into one step. Claude reads the issue, checks the linked PR history, and starts with full context. At the end of a coding session, Claude can update the issue status and add implementation notes without you touching Linear's UI.

Setup: Uses OAuth authentication. Run:

claude mcp add linear -- npx -y @anthropic/mcp-linear

Honest tradeoffs: Only useful if your team uses Linear. If you're on Jira or Asana, this isn't relevant (though Atlassian has their own MCP server for Jira).

Starter Stacks: What to Install Based on Your Role

You don't need all 12. Too many MCP servers consume context tokens, slow down startup, and give Claude more tools than it can use effectively. Pick 2-4 that match your daily work.

Full-Stack Developer

• GitHub (non-negotiable)
• Context7 (keeps your code suggestions current)
• Playwright (E2E testing)
• PostgreSQL (if you work with databases)

Frontend Developer

• GitHub
• Context7
• Figma (if your team uses it)
• Playwright (component testing in-browser)

DevOps / Platform Engineer

• GitHub
• Sentry (error triage)
• File System (cross-directory config management)
• Sequential Thinking (architecture and incident analysis)

Content Creator / Marketer

• Brave Search (research)
• Replicate (image generation)
• Notion (if your content calendar lives there)

Solo Builder / Indie Hacker

• GitHub
• Context7
• Brave Search
• Replicate (if you need visuals)

How Many MCP Servers Should You Actually Install?

The honest answer: fewer than you think.

Claude Code's MCP Tool Search feature added lazy loading, which reduces context usage by up to 95%. That's a significant improvement. But more servers still means more tools Claude needs to consider for each task, and tool selection accuracy tends to drop as the number of available tools increases.

In practice, 3-5 active MCP servers is the sweet spot for most developers. Enough to cover your core workflows. Few enough that Claude doesn't waste time deliberating which tool to use.

Start with two from Tier 1. Add one or two from Tier 2 or 3 based on your specific needs. If you find yourself constantly thinking "I wish Claude could access X," that's when you add another server. Not before.

Setting Up Your First MCP Server

If you haven't installed an MCP server before, here's the 2-minute version.

Step 1: Pick a server. For your first one, we recommend GitHub or Context7. Both have simple setup and immediate payoff.

Step 2: Run the add command.

claude mcp add context7 -- npx -y @upstash/context7-mcp@latest

Step 3: Verify it's working.

claude mcp list

You should see your server listed with its tools. If it shows 0 tools, something went wrong — check that Node.js is installed and the npx command runs successfully outside of Claude.

Step 4: Test it. Start a Claude Code session and try a task that uses the server. With Context7 installed, ask Claude to check the latest API for a library you use. It should fetch current docs rather than relying on training data.

For more detailed configuration — including project-scoped vs. user-scoped servers, environment variables, and troubleshooting — see the official Claude Code MCP documentation.

A Note on Security

MCP servers execute code on your machine and can access your data. A few things to keep in mind:

Scope permissions tightly. GitHub tokens should have minimal required scopes. Database connections should be read-only unless you specifically need write access. File system access should be limited to relevant directories.

Audit what you install. The MCP ecosystem doesn't have a security review process. Before installing a community-built server, check the source code. Look at the repo's recent commits, open issues, and contributor activity. Security researchers found hundreds of MCP servers with dangerous defaults in 2025 — the ecosystem has improved, but it's still your responsibility to vet what runs on your machine.

Keep servers updated. MCP servers get patched like any software. Run npx -y @package/latest to get the newest version, or pin specific versions if stability matters more than having the latest features.

FAQ

What is an MCP server?

An MCP (Model Context Protocol) server is a program that gives AI assistants like Claude Code access to external tools and data sources. It follows an open standard, so the same server works across Claude Code, VS Code, Cursor, and other compatible tools. Think of it as a plugin that extends what Claude can do — access your GitHub, query your database, search the web, or automate a browser.

Do MCP servers cost money?

The servers themselves are almost all free and open source. Some require API keys for the services they connect to — Brave Search has a free tier, Replicate charges per API call, GitHub is free for public repos. The MCP server is the bridge; the underlying service may or may not charge.

Can MCP servers break my code or data?

They can if given write permissions. An MCP server with write access to your database or repository can modify data based on Claude's instructions. Always start with read-only access for any server that connects to production systems or important data. Expand permissions only when you're confident in the workflow.

How are MCP servers different from Agent Skills?

MCP servers extend what Claude can access — your tools, databases, APIs, and services. Agent Skills extend what Claude knows how to do — domain expertise, frameworks, workflows, and decision logic. They're complementary. An MCP server connects Claude to your PostgreSQL database. A skill teaches Claude to write queries following your team's conventions and flag potential performance issues. The most effective setups use both.

Do MCP servers work with tools other than Claude Code?

Yes. The Model Context Protocol is an open standard. MCP servers work with Claude Code, Claude Desktop, VS Code with Copilot, Cursor, and other tools that support the protocol. A server you set up for Claude Code can typically be used with other MCP-compatible tools with minimal reconfiguration.

Why not just install every MCP server?

Three reasons. First, more tools means more context consumed per interaction, even with lazy loading. Second, Claude's tool selection gets less precise as options increase — with 50 tools available, it sometimes picks the wrong one. Third, each server is another piece of software running on your machine, consuming resources and expanding your attack surface. Install what you use. Skip what you don't.

---

Related reading:

• What Are Agent Skills? The Complete Guide — How skills and MCP servers work together
• Official Claude Code MCP Documentation — Full setup and configuration reference
• awesome-mcp-servers on GitHub — Community-curated list of MCP servers