Best MCP Servers for Claude Code in 2026: The Complete Setup Guide

Claude Code is the most powerful MCP client available today. Unlike Claude Desktop or Cursor, it supports three scope levels, acts as both MCP client and server, lazy-loads tools via Tool Search, and lets you teach it workflows through skills files. If you are using Claude Code without MCP servers, you are leaving most of its capability on the table.

Here are the 10 MCP servers worth installing, with the exact claude mcp add commands, and the Claude Code-specific features that make the setup different from every other client.

What Makes Claude Code Different as an MCP Client

Before the server list, three features that change how you should think about MCP in Claude Code versus Cursor or Claude Desktop.

Tool Search (lazy loading). Claude Code does not load every tool from every connected MCP server into context at once. It uses Tool Search to discover tools on demand, only pulling in the schema for tools it actually needs. This reduces context usage by roughly 95% compared to clients that dump all tool definitions upfront. The practical effect: you can connect 10+ MCP servers without drowning Claude in tool descriptions.

Three scope levels. Every MCP server you add has a scope:

• --scope user, Available globally, across all projects. Use for servers you always want (GitHub, Brave Search, Toolradar).
• --scope local, Per-user, per-project. Default scope. Good for project-specific servers (your database, Sentry for this project).
• --scope project, Writes to .mcp.json in the project root. Shared with your team via git. Use for servers the whole team needs.

Skills (.md files). You can create .claude/ files that teach Claude Code how to use your MCP tools in specific workflows. A skill file like .claude/deploy.md can say: "When I ask you to deploy, use the GitHub MCP to create a PR, then use Linear MCP to move the ticket to In Review." This turns MCP tools into composable workflows, something no other client supports.

What's New in MCP for Claude Code (Q1-Q2 2026)

The MCP ecosystem moved fast since launch. The four changes worth tracking:

1. Tool Search lazy-loading became default in Claude Code 4.x, connecting 10+ servers no longer floods context. Earlier versions loaded every tool definition upfront, which capped practical setups at 3-4 servers.
2. HTTP transport went mainstream. Most vendor-hosted servers (Sentry, Linear, Figma, Supabase, Notion) now ship as --transport http URLs with browser-based auth. Stdio servers still dominate for local-only servers (Filesystem, Playwright).
3. Skills (.md files) replaced the old slash-command-only flow. Composing multi-server workflows in .claude/ files is now the recommended pattern over piping commands manually.
4. First-party servers caught up to community ones. Notion, Stripe, Cloudflare, Slack, and Vercel all shipped official MCP servers in 2026, community forks are mostly obsolete now.

If you set up MCP servers before March 2026 and have not revisited the config, the official servers are usually faster, better-authenticated, and better-maintained than what you have.

The 10 Best MCP Servers for Claude Code

Tier 1: Install These Now

1. GitHub MCP Server

What it does: Full repository management, PRs, issues, code search, branches, commits, without leaving the terminal.

Claude Code already has deep git integration, but the GitHub MCP server adds the API layer: creating PRs with descriptions, searching issues across repos, reviewing PR comments, and managing labels. The combination of Claude Code's native file access and GitHub MCP's API access is the strongest code workflow available.

claude mcp add github --scope user --transport http https://api.githubcopilot.com/mcp/ \
  -e GITHUB_TOKEN=ghp_your_token

Or run it locally via Docker:

claude mcp add github --scope user \
  -- docker run -i --rm \
  -e GITHUB_PERSONAL_ACCESS_TOKEN=ghp_your_token \
  ghcr.io/github/github-mcp-server

Create a fine-grained personal access token at github.com/settings/tokens. Grant only the repos and permissions you need.

2. Context7 MCP

What it does: Fetches version-specific documentation for any library at call time.

Claude Code's training data is months old. When you are working with Next.js 15, Prisma 6, or Tailwind 4, Claude will generate code against outdated APIs. Context7 resolves the current documentation for any library and injects it into the conversation. Essential for any codebase using fast-moving dependencies.

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

No API key needed. Works immediately.

3. Brave Search MCP

What it does: Real-time web search using Brave's independent index.

Claude Code cannot access the internet by default. Brave Search gives it live web results, current documentation, recent release notes, Stack Overflow answers, security advisories. The most useful general-purpose MCP server for filling knowledge gaps.

claude mcp add brave-search --scope user \
  -- npx -y @brave/brave-search-mcp-server

Set your API key as an environment variable: export BRAVE_API_KEY=your_key in your shell profile. Free tier: 2,000 queries/month at brave.com/search/api.

4. Toolradar MCP

What it does: Search, compare, and get pricing for 8,400+ software tools with verified data.

"What's the best headless CMS for a Next.js project?" or "Compare Supabase vs Firebase pricing", Claude answers from real data instead of hallucinating stale information. Editorial scores, weekly-verified pricing, G2/Capterra ratings, and AI-identified alternatives.

claude mcp add toolradar --scope user \
  -- npx -y toolradar-mcp

Set your API key: export TOOLRADAR_API_KEY=tr_live_your_key. Free key (100 calls/day): toolradar.com/dashboard/api-keys. Full docs.

Tier 2: Add Based on Your Stack

5. Playwright MCP

What it does: Browser automation, navigate pages, click elements, fill forms, take screenshots, run E2E tests.

Claude Code can write Playwright tests natively. But with the Playwright MCP server, it can run them too, execute tests, take screenshots of failures, and iterate on fixes without you switching to a browser. The feedback loop tightens dramatically.

claude mcp add playwright --scope local \
  -- npx -y @playwright/mcp --headless

No API key. Requires browsers installed: npx playwright install chromium.

6. Sentry MCP

What it does: Pull error context, stack traces, breadcrumbs, affected users, directly into Claude Code.

"What's the top error this week?" returns the actual stack trace, not a guess. Claude reads the full error context and suggests fixes against your codebase. Use --scope local so it connects to the right project.

claude mcp add sentry --scope local \
  --transport http https://mcp.sentry.dev/mcp

Authenticate via the browser flow on first use.

7. Figma MCP

What it does: Exposes your Figma selection's live structure, hierarchy, auto-layout, spacing tokens, component refs.

Claude Code generates code against the real design, not a screenshot. Paste a Figma link, and Claude reads the component tree and produces matching code with accurate spacing, colors, and layout. Especially strong for Tailwind-based projects.

claude mcp add figma --scope user \
  --transport http https://mcp.figma.com/mcp

8. Supabase MCP

What it does: Manage your Supabase project, database, auth, storage, edge functions, from Claude Code.

If your stack includes Supabase, this replaces the dashboard. Create tables, manage RLS policies, debug auth issues, query data, all from the terminal conversation.

claude mcp add supabase --scope local \
  --transport http https://mcp.supabase.com/mcp

9. Linear MCP

What it does: Read, create, and update issues. Manage sprints. Search across projects.

"Create a bug ticket for the auth redirect issue, assign to me, add to the current sprint", done in one message. When combined with the GitHub MCP server, Claude Code can create the fix PR and update the Linear ticket in a single workflow.

claude mcp add linear --scope user \
  --transport http https://mcp.linear.app/mcp

10. Filesystem MCP Server

What it does: Secure file operations with sandboxed directory access.

Claude Code already has native file access within your project. The Filesystem MCP server adds value when you need Claude to access files outside the project directory, shared config directories, monorepo sibling packages, or data directories. Specify exactly which directories to expose.

claude mcp add filesystem --scope local \
  -- npx -y @modelcontextprotocol/server-filesystem /path/to/allowed/dir

No API key needed. Only the directories you specify are accessible.

My Recommended Setup

Copy-paste these commands to get a strong baseline. Run them once, they persist across sessions.

# Global servers (available in every project)
claude mcp add github --scope user --transport http https://api.githubcopilot.com/mcp/ -e GITHUB_TOKEN=ghp_your_token
claude mcp add context7 --scope user -- npx -y @upstash/context7-mcp
claude mcp add brave-search --scope user -- npx -y @brave/brave-search-mcp-server
claude mcp add toolradar --scope user -- npx -y toolradar-mcp

# Project-specific (run inside your project directory)
claude mcp add playwright --scope local -- npx -y @playwright/mcp --headless
claude mcp add sentry --scope local --transport http https://mcp.sentry.dev/mcp

Set environment variables in your shell profile (~/.zshrc or ~/.bashrc):

export BRAVE_API_KEY="your_brave_key"
export TOOLRADAR_API_KEY="tr_live_your_key"

Verify everything loaded:

claude mcp list

Claude Code as an MCP Server

This is the feature no other client has. Claude Code can act as an MCP server itself, meaning other tools and agents can call it.

claude mcp serve

This exposes Claude Code's capabilities (file editing, bash execution, code analysis) to any MCP client. Use cases:

• Orchestration. A parent agent delegates coding tasks to Claude Code via MCP.
• CI/CD integration. A pipeline calls Claude Code to review PRs, generate changelogs, or fix lint errors.
• Multi-agent systems. One agent handles planning, another (Claude Code) handles implementation.

This turns Claude Code from a development tool into a development service that other systems can invoke programmatically.

Sharing MCP Config with Your Team

Use --scope project to write server definitions to .mcp.json in your project root. This file gets committed to git so everyone on the team uses the same MCP setup.

claude mcp add context7 --scope project -- npx -y @upstash/context7-mcp
claude mcp add playwright --scope project -- npx -y @playwright/mcp --headless

This creates .mcp.json:

{
  "mcpServers": {
    "context7": {
      "command": "npx",
      "args": ["-y", "@upstash/context7-mcp"]
    },
    "playwright": {
      "command": "npx",
      "args": ["-y", "@playwright/mcp", "--headless"]
    }
  }
}

Keep API keys out of this file. Each developer sets their own keys via environment variables or --scope local.

Writing Skills for MCP Workflows

Skills are .md files in your project's .claude/ directory that teach Claude Code compound workflows using MCP tools. Example:

.claude/commands/triage-bug.md

When I say "triage [error-id]":
1. Use Sentry MCP to fetch the full error context for the given ID
2. Find the relevant source file in this codebase
3. Propose a fix with explanation
4. Use Linear MCP to create a bug ticket with the error summary and proposed fix
5. Ask me if I want to create a PR with the fix

Now typing /triage SENTRY-ABC-123 executes the full workflow across multiple MCP servers in sequence. This is where Claude Code's MCP support becomes qualitatively different from other clients, you are building reusable, composable automation, not just connecting tools.

Security

MCP servers run code on your machine and call external APIs with your credentials. Three rules:

1. Audit before installing. Read the source of any MCP server before running it. Prefer official servers (published by the tool vendor) over community forks.
2. Principle of least privilege. Use read-only database users, fine-grained API tokens, and scoped file access. Never give a server more access than it needs.
3. Scope carefully. Global (--scope user) servers are available everywhere, only promote a server to global if you trust it completely.

For a deeper treatment, see our MCP server security best practices guide.

FAQ, Claude Code MCP Setup

What is the difference between --scope user, --scope local, and --scope project?
--scope user is global across every project on your machine, use for servers you always want (GitHub, Brave Search, Toolradar). --scope local is the default and writes to a per-project, per-user config, use for project-specific servers like a database or Sentry project. --scope project writes to .mcp.json in the repo root and gets committed to git so the whole team shares it. Pick project only for servers your teammates need; pick local for anything with personal credentials.

Do I need to restart Claude Code after adding an MCP server?
No. New servers are picked up on the next message. Run claude mcp list to confirm the server registered. If a server fails to load, the error appears in claude mcp doctor.

How many MCP servers can I install at once?
Practically unlimited since Tool Search lazy-loads schemas. The bottleneck is no longer context, it is which credentials you trust to expose. We have run setups with 15+ servers without measurable performance hit.

What is the difference between MCP servers in Claude Code vs Claude Desktop?
Claude Code supports three scope levels (user/local/project), runs in your terminal with full filesystem and shell access, and can act as an MCP server itself via claude mcp serve. Claude Desktop is a single global config, sandboxed (no shell), and cannot serve. The same MCP server packages usually work in both, but Claude Code's scope model is what makes team workflows possible.

Why does my GitHub MCP server fail with 401 / unauthorized?
The token expired, the token does not have the right scopes, or it is the wrong format. Use a fine-grained personal access token (not a classic PAT) with explicit scope on the repos you want, issue/PR/contents read-write at minimum. Test the token directly: curl -H "Authorization: Bearer $GITHUB_TOKEN" https://api.github.com/user. If that 401s, the token itself is invalid.

Can I use MCP servers without giving them my API keys?
For HTTP servers with browser-based auth (Sentry, Linear, Figma, Supabase) yes, Claude Code triggers an OAuth flow on first call and stores the token securely. For stdio servers (GitHub via Docker, Brave Search, Toolradar) you provide the API key as an environment variable or -e flag. Either way, keys never end up in your repo as long as you keep them out of .mcp.json.

Are MCP servers safe? What if a server gets compromised?
MCP servers run code on your machine and call external APIs with your credentials. Three controls reduce blast radius: (1) prefer official servers (vendor-published) over community forks; (2) use read-only credentials and least-privilege API tokens; (3) keep dangerous servers (filesystem write, shell exec) at --scope local so they only run when you are inside that project. A compromised server with read-only API access leaks data; a compromised server with write access can corrupt it.

How do I update an MCP server to a newer version?
For npx-based stdio servers (-- npx -y package-name), they auto-update on each invocation, npx pulls the latest each time. For Docker-based servers, run docker pull <image>. For HTTP servers, the vendor handles updates server-side. Run claude mcp doctor periodically to spot deprecated config formats.

My MCP server is loaded but Claude is not using it. Why?
Three causes: (1) the user prompt does not signal the right tool, try mentioning the tool by name ("use the Linear MCP to..."); (2) Tool Search has not surfaced the relevant tools yet, run claude mcp doctor to confirm tool definitions are reaching the model; (3) the server returned errors silently, check the server's logs (each one logs to a different place, usually documented in the server's README).

Troubleshooting Common MCP Issues

claude mcp list shows the server but tools never appear in conversation.
Run claude mcp doctor, it will show whether the server's tool definitions reached the runtime. If they did not, the server probably crashed at startup. Test the server directly: copy the command and args from claude mcp list --json and run them in a terminal. Most stdio servers will print a JSON-RPC handshake on stdin/stdout. If that hangs or errors, the server is misconfigured.

The server keeps reconnecting / dropping in the middle of a request.
For stdio servers, the most common cause is the server killing itself due to an uncaught exception. Check the server's logs. For HTTP servers, the vendor's session expired, re-run the auth flow with claude mcp auth <server-name> if available, or remove and re-add the server.

Tool descriptions show up in Claude's reasoning but it never actually calls them.
Either the tool's input schema is too strict (Claude refuses to call it because it cannot satisfy required fields) or the tool description is too vague to match the user request. Inspect the tool with claude mcp tools <server-name> and confirm the descriptions read like instructions, not API doc dumps.

Permission denied when adding a --scope user server.
The user-scope config lives at ~/.claude/mcp.json (or %APPDATA%\Claude\mcp.json on Windows). Confirm the directory is writable. On macOS, this can fail if the directory was created by a different user (e.g., during a sudo install).

The server works locally but breaks for teammates after committing .mcp.json.
You probably committed env-var-bearing keys, or the server depends on a global tool (Docker, npx, browsers) that teammates do not have. Document required prerequisites in the project README, and never put secrets in .mcp.json, every developer should set their own keys via shell environment.

Full list of 25 servers for any client: Best MCP Servers in 2026

Setting up MCP for Claude Desktop instead? Claude Desktop MCP Setup Guide

Want your AI agent to query Toolradar? Set up Toolradar MCP