Top 10 MCP Servers for Claude Code in 2026 — Tested, Ranked, Compared

The Model Context Protocol (MCP) registry crossed 400 servers in early 2026. That sounds like abundance until you start installing them and discover that maybe a quarter still work, half have unclear documentation, and a meaningful fraction were one-shot demos that nobody maintains. Picking the right MCP servers for Claude Code is now a real skill.

This guide cuts through the noise. We've installed and tested every MCP server on this list inside real Claude Code workflows for at least a week. Each entry includes the install command, the one-sentence pitch, the workflow it actually unlocks, and (when relevant) what to use instead. If you only install what's on this page, you'll have a Claude Code setup that handles 90% of real engineering work.

For configuration basics — where MCP server definitions live, how Claude Code discovers them — see our .claude/ folder complete guide.

Three criteria, applied ruthlessly:

1. Used in real workflows in 2026 — not just installed once for a demo. Public usage telemetry, GitHub stars trending up, recent commit activity.
2. Solves a class of problems Claude Code can't handle natively — if a built-in tool already covers it, the MCP doesn't earn a slot.
3. Doesn't break frequently — failures during normal use, dependency drift, brittle setup all push a server off the list.

Servers ranked roughly by frequency of use. Your mileage will vary depending on whether you're doing frontend, backend, data, or content work — see the workflow stacks section for tailored recommendations.

1. Filesystem MCP (official reference server)

Pitch: cross-project file access without changing Claude Code's working directory. Read files in /etc/, write to ~/Desktop, copy between sibling repos.

Install: claude mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem /Users/you/projects

Why it earns slot 1: every other MCP server can do something narrow. Filesystem is the universal bridge. If you only install one MCP, install this one.

2. GitHub MCP

Pitch: search code across all your repos, open issues and PRs from the prompt, read PR diffs. Claude can finally answer "where in our 30 repos do we use this function?"

Install: claude mcp add github -- npx -y @modelcontextprotocol/server-github (set GITHUB_PERSONAL_ACCESS_TOKEN env var)

Replaces: hand-rolled gh CLI scripts. The MCP version is faster and Claude can chain calls (search → diff → comment).

3. Playwright MCP

Pitch: real browser automation. Take screenshots, fill forms, scrape JavaScript-rendered content, run E2E tests with Claude orchestrating the steps.

Install: claude mcp add playwright -- npx -y @playwright/mcp

Real use case: scraping API documentation that requires JavaScript to render, then having Claude generate a typed client. Saves an hour every time.

4. Tavily MCP

Pitch: web search with structured results, much higher quality than ad-hoc web fetching. Critical for anything time-sensitive (libraries released in 2026, recent CVEs).

Install: claude mcp add tavily -- npx -y tavily-mcp (set TAVILY_API_KEY)

Why prefer over plain WebSearch: Tavily returns extracted content with relevance scores. WebSearch returns links you have to fetch separately. Token efficiency is roughly 3× better.

5. Context7 MCP

Pitch: real-time documentation for any library on npm/PyPI/etc. Claude pulls the actual current docs instead of hallucinating from training cutoff knowledge.

Install: claude mcp add context7 -- npx -y @upstash/context7-mcp

When it matters most: any library that has shipped breaking changes since Claude's training cutoff. react-router, tanstack-query, tailwindcss all changed APIs in 2025-2026; Context7 keeps Claude honest.

6. Sequential Thinking MCP

Pitch: explicit step-by-step reasoning for complex problems. Claude breaks the work into numbered steps, can revise, and the trace is visible in the conversation.

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

Best for: architecture decisions, root-cause debugging, multi-component refactors. Skip it for trivial tasks — the overhead is real.

7. Postgres / SQLite MCP

Pitch: query your database directly from Claude. Schema introspection, read queries, write queries (gated by permissions). Replaces the dance of "show me the table, OK now query it."

Install: claude mcp add postgres -- npx -y @modelcontextprotocol/server-postgres "postgresql://..."

Safety note: configure read-only credentials for the MCP. Never give it production write access.

8. Memory MCP

Pitch: persistent knowledge graph across Claude Code sessions. Remembers entities (your team, your codebase modules) and relations between them.

Install: claude mcp add memory -- npx -y @modelcontextprotocol/server-memory

Honest assessment: the value depends entirely on whether you build the habit of saving useful facts. If you don't, it's a slot wasted. If you do, it compounds over months.

9. Apidog MCP

Pitch: API design and testing. Define endpoints, run requests, share collections. Replaces Postman for AI-driven API workflows.

Install: claude mcp add apidog -- npx -y @apidog/mcp

When it earns its slot: backend dev who frequently writes new endpoints. The "design endpoint, test endpoint, generate typed client" loop becomes one prompt.

10. Sentry MCP

Pitch: pull error reports and performance traces directly into the Claude session. "Why is this endpoint slow in prod?" gets answered with real data instead of speculation.

Install: claude mcp add sentry -- npx -y @sentry/mcp-server (configure org + project)

Honest take: only earns its slot if you actually use Sentry. Replace with the equivalent for Datadog / Honeycomb / New Relic if that's your stack.

Five servers that didn't quite make the top 10 but are worth knowing about for specific workflows:

• Slack MCP — read channels, post messages, search history. Earns a slot if your team lives in Slack and you want to summarize threads or post status updates from Claude.
• Notion MCP — query and update Notion pages. Useful if Notion is your team's source of truth for specs and runbooks.
• Linear MCP — create and update issues. Pairs well with GitHub MCP for "find the bug in code, file the issue, link both."
• Magic MCP (21st.dev) — generate UI components from natural-language descriptions. Useful for rapid frontend prototyping; less useful in mature codebases with established component libraries.
• Stripe MCP — query Stripe data (customers, subscriptions, charges) for support and analytics. Niche but invaluable if your work touches billing.

What we left off intentionally: anything with fewer than 100 GitHub stars, anything not updated since Q1 2025, and any "AI agent framework" that's really an MCP wrapper around an LLM call (those add latency without unlocking new capability).

The right stack depends on what you do. Three concrete recommendations:

Stack A: Frontend Developer

• Filesystem (slot 1)
• GitHub (slot 2)
• Playwright (slot 3) — for visual testing
• Context7 (slot 5) — for current React/Tailwind/Vite docs
• Magic MCP (honorable mention) — for component generation

Skip: Postgres, Sentry, Apidog (unless you also do backend).

Stack B: Backend Developer

• Filesystem
• GitHub
• Postgres / SQLite
• Apidog
• Sentry (or Datadog/Honeycomb equivalent)
• Sequential Thinking (slot 6) — for architecture decisions

Skip: Playwright, Magic.

Stack C: Technical Writer / Content Creator

• Filesystem
• Tavily (slot 4) — for research
• Context7 (slot 5) — for accurate API references in tutorials
• Memory (slot 8) — for tracking style guide decisions across pieces
• Notion (honorable mention) — if your CMS is Notion

Skip: GitHub (unless documenting open source), Postgres, Apidog.

Don't install all 10 at once

Each MCP server adds latency to Claude Code session startup and memory overhead. Five is a sweet spot for most users; ten is the point at which startup feels noticeably slow. If you go above ten, audit which ones you've actually used in the last week and remove the rest.

Where do MCP servers get configured in Claude Code?

claude mcp add writes to ~/.claude.json by default. For project-scoped servers (where the team should share the config), use claude mcp add --scope project, which writes to .mcp.json at the repo root.

How do I check which MCP servers are currently active?

claude mcp list shows all configured servers. claude mcp get <name> shows the full config for one. Inside a session, /mcp lists active connections.

Can I write my own MCP server?

Yes — the protocol is open and there are SDKs for TypeScript, Python, and Rust. Start with the official template at github.com/modelcontextprotocol/create-server. Plan on a few hours for a basic server, longer for one that handles state.

Do MCP servers work in Claude Code's CLI mode and the desktop app?

Both. Configurations roam between the CLI, the desktop app, and IDE extensions because they all read the same ~/.claude.json.

What's the difference between an MCP server and a custom subagent?

An MCP server adds tools Claude can call. A subagent is a specialized Claude persona with its own context window. They compose: a custom subagent can use any MCP-installed tool. See our .claude/ folder guide for subagent details.

How do I stop an MCP server I no longer use?

claude mcp remove <name>. Removing it from ~/.claude.json manually also works. Restart any active Claude Code sessions for the change to take effect.

Are MCP servers a security risk?

They have whatever permissions you give them and run as your user. Audit before installing — read the source, check the maintainer, look at recent issues. Treat unknown MCP servers like any unknown shell tool: npx-style execution gives them full access to your machine.

Related: .claude/ folder complete guide · Claude Code Hooks practical guide · Claude Buddy Checker