You install Claude Code, run it for a week, and the same question hits every developer I know: which MCP servers actually deserve a slot in my config? The official directory in 2026 has over 200 entries, and a lot of them are someone's weekend hack that hasn't been updated in eight months. I've cycled through dozens of them. This is the shortlist that earned a permanent home.
If you only do one thing after reading this: install Filesystem and Playwright. Those two cover most of the value of the entire MCP ecosystem and they'll change how you use Claude Code immediately.
The starter pack — install these first
1. Filesystem
The single most useful MCP server, and the one Anthropic should bundle by default. Claude Code can already read files in your project root — that's its bread and butter. What Filesystem unlocks is everything outside that root: your ~/Documents, scratch directories, notes, downloaded PDFs, sibling repos. I use it constantly to pull context from neighbor projects without having to cd around.
Install: claude mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem ~/Documents ~/GitHub
Gotcha: scope it tightly. Don't grant access to ~/ wholesale — Claude will happily index your entire home directory and burn tokens on cached browser data.
2. Playwright
Playwright turns Claude Code into a browser. Open a page, click around, take screenshots, read the DOM, fill forms, watch network requests. The use cases just keep showing up: debugging your own deploy ("does this URL actually work?"), scraping a one-off page, taking a screenshot for a bug report, testing a flow end-to-end. Before Playwright I'd kick Claude over to a terminal to use curl; now it just opens the page.
Install: claude mcp add playwright -- npx -y @modelcontextprotocol/server-playwright
Gotcha: Playwright downloads a Chromium binary on first run. ~150MB. Be patient.
The team-workflow tier — install if you work with others
3. GitHub
Lets Claude create PRs, read issues, comment, manage labels, kick off Actions. The official server is reliable and the auth flow is clean (it uses your gh CLI token). If you live in GitHub, this is the one. I use it for "look at PR #142 and summarize what changed," then "open a PR with these changes to my-repo."
Install: claude mcp add github -- npx -y @modelcontextprotocol/server-github
4. Linear
Same idea, different ticketing system. Create tickets from Claude, update status, move them across boards, query backlogs. The Linear server is particularly good because Linear's API is so well-designed — Claude rarely fumbles it. Use for "create a P1 ticket assigned to me, label backend, with these acceptance criteria."
Install: claude mcp add linear -- npx -y @modelcontextprotocol/server-linear
5. Slack
Send messages, read channels, search threads. The killer use case is "summarize the eng-incidents channel from the last 24 hours" before you scroll through 400 messages yourself. Also "post a deploy summary to #releases." Read-only mode is the safer default — full write access is powerful but requires more careful permission scoping.
Install: claude mcp add slack -- npx -y @modelcontextprotocol/server-slack
The data tier — install if you query databases
6. Postgres
Read-only Postgres access. Hand Claude a connection string and it can query your dev or read-replica database. I use this constantly during debugging: "how many rows in the orders table have null user_id?" beats writing SQL manually. The write side is intentionally locked off in the official server, which is the right call.
Install: claude mcp add postgres -- npx -y @modelcontextprotocol/server-postgres "postgresql://localhost/mydb"
7. Sentry
Read errors, search issues, get stack traces. Sentry has so much context that's painful to dig through manually — Claude can summarize ten error events into "looks like a missing null check on user.email when the new SSO path is hit." This one earns its keep when you have an active incident.
Install: claude mcp add sentry -- npx -y @modelcontextprotocol/server-sentry --auth-token
The "I have a homelab" tier
8. SQLite
Pointed at any SQLite file, lets Claude query it. Sounds niche, but if you run any local data project — note databases, downloaded datasets, browser-history exports — this is the fastest way to ask questions of it.
9. Memory
A persistent key-value store Claude can write to and read across sessions. The official server is simple; the value depends on how disciplined you are about using it. I save common prompts, code style preferences, and project-specific quirks. It's like giving Claude a sticky-note pad.
10. Fetch
Make HTTP requests from Claude. Lighter weight than Playwright when you just want JSON from an API. Good for testing your own webhooks, querying public APIs ("what's the current Bitcoin price"), and quick integration prototyping.
The ones I'd skip even though they get recommended
Browser-Use — superseded by Playwright. Pick Playwright.
Time / DateTime servers — Claude already knows the current date. Adding a server for this is pure overhead.
Wikipedia / Wolfram — neat demos but you'll use them once and forget. Claude's built-in knowledge plus web search handles 99% of factual queries.
Random "browse-this-app" servers from individual creators — most don't update, some have auth issues. Check last commit date and star count before adding.
Performance reality check
Each MCP server adds about 1–3 seconds to Claude Code's startup time and a tiny latency tax on each request. Five servers is fine. Ten is noticeable. Twenty makes the launcher feel sluggish and the agent's tool-selection step slower because there's more to consider. I run six on my main config and that's the sweet spot.
The verdict — install in this order
Day one: Filesystem, Playwright. These two alone justify the entire MCP system existing.
Day two, if you use them: GitHub, Linear or Jira, Slack. Wire up the tools you spend the most time in.
When you next debug something: Postgres (for the dev DB), Sentry (for production errors).
When you find a use case: Memory, Fetch, SQLite. Don't install these preemptively — wait until you have a real reason.
Resist the temptation to load up. The point of MCP isn't to have every tool installed — it's to have the right ones ready when you need them.
FAQ
What are MCP servers in Claude Code?
Small plugins that give Claude Code new tools — read databases, open browsers, create tickets, send Slack messages. Each runs as a local process and exposes its tools to Claude.
How do I install one?
One command in most cases: claude mcp add <name> -- npx -y @modelcontextprotocol/server-<name>. Claude Code's MCP wizard handles permissions and reload.
Do they slow Claude Code down?
Each server adds 1–3 seconds to startup. Five is fine. Twenty is noticeable. Install what you use; remove what you don't.
Which one should I install first?
Filesystem and Playwright. They cover the majority of the ecosystem's value. Everything else is situational.
Are MCP servers safe?
Official Anthropic and reference servers are safe. Third-party ones deserve the same scrutiny as any npm package — check the repo, recent activity, and what permissions they request.