The 10 Best MCP Servers for Claude Code in 2026

If you're using Claude Code without MCP servers, you're missing maybe 40% of what it can do. Out of the box, Claude Code is excellent at writing code. With the right MCP servers wired up, it becomes an assistant that can read your database, manage your GitHub issues, query your error tracker, deploy to your hosts, and pull design specs from Figma — without you switching tabs once.

I've spent the last few months testing essentially every MCP server I could find. Most are fine. A handful are genuinely transformative. Here are the 10 that earned a permanent spot in my Claude Code config — and the case for each.

How I picked these #

A "best of" list is only useful if you know the criteria. Here's what I weighted:

• Real productivity gain — does it save meaningful time, not just look cool in a demo
• Reliability — does it work consistently, or does it break twice a week
• Setup effort — bonus points for "install and go," penalty for "configure 12 environment variables"
• Generality — useful across many projects, not just one specific stack
• Maintenance — actively developed, not abandoned six months ago

These aren't ranked strictly best-to-worst. They're listed in roughly the order I'd install them as a new Claude Code user.

1. Filesystem #

The starting point. The Filesystem server lets Claude Code read, write, and search files in directories you grant access to. If you're using Claude Code to edit code, you need this — there's no good substitute.

What makes it the first install: it's official (Anthropic-built), zero API keys, zero config beyond pointing it at a folder. Claude can read your files, search across them, and write new ones. Most other servers assume you have this one running.

Best for: every Claude Code user, every project.

2. GitHub #

The GitHub MCP server turns Claude Code into a teammate who actually understands your repo. Issues, pull requests, code search across all your repos, file content fetching, commit history — all addressable through natural language.

The killer feature: Claude can search your entire org's code, not just the repo you have open. "Find all the places we initialize the Postgres pool across our microservices" becomes a one-line ask. You'll never paste GitHub URLs again.

Setup takes 2 minutes (generate a personal access token with repo scope, paste into config).

Best for: anyone working in a team or with multiple repos.

3. Postgres #

If you're building anything with a database, install Postgres immediately. It's read-only by default, which is the right tradeoff: you get all the benefits (schema introspection, real query results, debugging help) without the risk of Claude accidentally dropping your users table.

The actual workflow that makes this transformative: instead of writing SQL, switching to your DB client, running the query, eyeballing the results, switching back, and pasting into Claude — you just ask Claude. It writes the query, runs it, reads the results, and explains them. The feedback loop collapses from minutes to seconds.

Best for: backend developers, data engineers, anyone whose work touches a Postgres database.

4. Sentry #

The Sentry MCP server lets Claude pull real production errors directly into your conversation. Not screenshots. Not pasted stack traces. Live data.

The use case that sold me: I'd see a Sentry alert come in. Open Claude Code. Ask "what's that error in checkout right now and what's the stack trace?" Claude pulls the issue, summarizes the stack, looks at the relevant files in my repo, and proposes a fix. Total time: maybe 90 seconds. Without the MCP server, the same task is a 10-minute context-switching dance.

Best for: anyone whose code runs in production with real users.

5. Linear #

If your team uses Linear (and a lot of fast-moving teams do), the Linear MCP server is essential. Claude can read your assigned issues, comment on tickets, create new issues from conversations, and reason about what's in flight.

The pattern I use most: at the start of a coding session, "what are my Linear issues for this sprint and which ones are blocked?" Claude lists them, I pick one, we work on it, Claude updates the ticket when we're done. Project management without leaving the IDE.

If you're on Jira instead, look at the Jira MCP server (community-built, less polished but functional).

Best for: teams using Linear; product engineers especially.

6. Fetch #

Deceptively simple, ridiculously useful. The Fetch server lets Claude grab any URL and return it as clean Markdown. It's the canonical "give Claude internet access" tool.

Why it earned a permanent spot: documentation. When I'm working with a library and Claude isn't sure of the latest API, "fetch the docs at [URL] and tell me how to do X" almost always nails it. Same for blog posts, GitHub READMEs, and standards documents. Claude with Fetch is way more accurate than Claude without.

Pairs perfectly with Firecrawl if you need bulk scraping or JS-rendered pages.

Best for: everyone. Install it, you'll find uses you didn't expect.

7. Memory #

The Memory server gives Claude long-term memory across sessions. It's a simple knowledge graph that stores facts you tell Claude to remember and retrieves them in future conversations.

I was skeptical at first — felt like overkill. Then I started using it for project context. "Remember: this codebase uses pnpm workspaces, the staging URL is X, our convention is to put types in /shared/types." Six weeks later Claude still remembers, and I never have to repeat myself when starting a new conversation.

It's not magic — you have to be explicit about what to remember. But the payoff is huge if you have ongoing projects.

Best for: anyone with long-running projects or recurring contexts.

8. Playwright #

The Playwright MCP server (Microsoft-built) lets Claude drive a real browser. Navigate to URLs, click elements, fill forms, take screenshots, evaluate JavaScript on the page. The same Playwright engine you use for tests, exposed through MCP.

Use cases that make this worth installing:

• Reproducing bug reports ("go to staging.app.com, log in as test@user, click Settings, screenshot what you see")
• Quick scraping when Fetch isn't enough (JS-heavy sites)
• Visual regression checks ("compare this screenshot to the design in Figma")
• Live API exploration through admin UIs

It's heavier than Fetch — it actually launches a browser — so don't install if you don't need it. But if you do browser work regularly, it's transformative.

Best for: QA engineers, full-stack developers, anyone debugging frontend issues.

9. Figma #

If you work with designers, the Figma MCP server (Framelink-built) is essential. Claude can read Figma files, frames, and components directly — meaning "implement this design" stops being a copy-paste-from-screenshot dance.

The workflow: designer drops a Figma link in Slack. You paste it into Claude Code with "implement this in our React codebase using our existing components." Claude reads the design specs (colors, spacing, typography, component hierarchy), maps them to your code conventions, and produces a draft. Most of it works. The parts that don't, you iterate on with the design still loaded.

Best for: frontend developers, design-engineering, anyone implementing UI.

10. Vercel (or your hosting MCP) #

The Vercel MCP server lets Claude inspect your deployments — see build logs, check production status, look at preview URLs, check environment variables. Combined with GitHub, this means Claude can deploy code, watch the deploy, and check for errors all in one flow.

If you're not on Vercel, the equivalents to look at are Cloudflare (great if you're on Workers/Pages) and AWS (more complex but powerful for AWS shops).

Best for: anyone who deploys regularly. The deploy-check-error loop is the most repetitive part of shipping software, and an MCP server flattens it.

Honorable mentions #

A few I almost included:

• Notion — if your team docs live in Notion, this is high-value
• Slack — useful for searching conversation history, but Claude reading Slack feels invasive to coworkers
• Stripe — niche but indispensable if you work on billing/payments
• Brave Search — solid alternative to Fetch for general web searches
• E2B Code Interpreter — sandboxed code execution; excellent for math/data tasks

Browse the full directory to find servers for your specific stack — there's almost certainly one already built for the tools you use.

Setting up MCP servers in Claude Code #

Claude Code's MCP config lives in a slightly different place than Claude Desktop:

~/.claude/claude_code_config.json

Or, run claude config in your terminal to open the config in your default editor.

The format is identical to Claude Desktop:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/your/projects"]
    },
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_yourtoken"
      }
    },
    "postgres": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-postgres", "postgresql://localhost/mydb"]
    }
  }
}

Restart Claude Code after editing. You should see the tools become available — typically a tools indicator in the status line, depending on your client.

A realistic install order for new users #

If 10 sounds like a lot to install at once (it is), here's the order I'd add them in. Each builds on the last:

Week 1: Filesystem, GitHub, Fetch (the foundation)

Week 2: Postgres + your hosting MCP (Vercel/Cloudflare/AWS) (your stack)

Week 3: Sentry, Linear (your operational tools)

Week 4: Memory, Figma, Playwright (the optional but useful set)

After a month, you'll have a Claude Code setup that's roughly 3x more capable than where you started — and you'll know exactly which servers earned their place because you saw the difference each one made.

What I'm watching for next #

The MCP ecosystem is growing fast. The categories I expect to see breakout servers in over the next few months:

• CI/CD inspection — read build logs, trigger reruns, debug pipelines (currently weak)
• Observability — beyond Sentry: Datadog, Grafana, full APM through MCP
• Browser session sharing — letting Claude pick up where you left off in a browser session, with auth carried over
• Native voice/audio — for voice-first workflows on mobile

If you've built one of these, submit it to the directory — we add new servers weekly and these are the categories I'm specifically curating for.

TL;DR install list #

In one block, here's the config to install all 10 (you'll need to add API keys/connection strings where required):

{
  "mcpServers": {
    "filesystem": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "/your/projects"] },
    "github": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"], "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..." }},
    "postgres": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-postgres", "postgresql://..."] },
    "sentry": { "command": "npx", "args": ["-y", "@sentry/mcp-server"], "env": { "SENTRY_AUTH_TOKEN": "..." }},
    "linear": { "command": "npx", "args": ["-y", "@linear/mcp-server"], "env": { "LINEAR_API_KEY": "..." }},
    "fetch": { "command": "uvx", "args": ["mcp-server-fetch"] },
    "memory": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-memory"] },
    "playwright": { "command": "npx", "args": ["-y", "@playwright/mcp"] },
    "figma": { "command": "npx", "args": ["-y", "figma-developer-mcp"], "env": { "FIGMA_API_KEY": "..." }},
    "vercel": { "command": "npx", "args": ["-y", "@vercel/mcp-server"], "env": { "VERCEL_TOKEN": "..." }}
  }
}

Drop that in your config, fill in the credentials, restart Claude Code. That's the modern Claude Code dev environment in 2026.

Next steps #

• New to MCP? Start with our installation guide
• Still deciding between MCP and function calling? Read our comparison
• Want to discover more servers? Browse the full directory
• Built one we should know about? Submit it

What I'd love to hear: which of these surprised you, and which one is missing from your stack? Email me at hello@mcpcatalog.dev — I read every one.