If you only connect a few MCP servers to Claude Code, make it these: Context7 for live library docs, Playwright for browser automation, and the official server for whatever platform you deploy on (Supabase, Vercel, Cloudflare, GitHub). Almost everything else is situational — and connecting servers you don't use makes Claude slower and more confused, not more powerful.

Here's the 2026 shortlist, what each server is actually good for, and the honest cases where you should skip one.

First, how to add a server

Three ways, depending on scope:

# Local (stdio) server, just for you
claude mcp add context7 -- npx -y @upstash/context7-mcp

# Remote server over HTTP (OAuth handled in-client)
claude mcp add --transport http vercel https://mcp.vercel.com

# Project-scoped: commit a .mcp.json so the whole team gets it

The big shift since 2025: the serious platforms now run hosted remote servers with OAuth — you authenticate once in the client and never manage an API key. Prefer the remote option when a vendor offers both. We covered why in MCP connectors are changing how enterprise AI agents connect to real systems.

Documentation and context

Context7 — pulls current, version-specific library documentation into context on demand. Two tools, near-zero setup, and it visibly cuts the "Claude wrote code against an API that changed two majors ago" class of bug. If you work with fast-moving frameworks, this is the single highest-value server on the list.

Fetch (from the official reference servers) — fetches a URL and converts it to markdown. Claude Code ships its own web fetch, so you mostly don't need this one — it earns its place in stripped-down or self-hosted setups.

Browser automation and testing

Playwright MCP — Microsoft's official server. Claude drives a real browser: click, fill, screenshot, read the accessibility tree, check console errors. This is how you get Claude to actually verify the UI it just built instead of declaring victory from the diff. Honest caveat: browser sessions are token-hungry, so connect it when you're doing frontend work, not permanently.

Your deploy platform (pick yours)

The pattern that works: connect the one or two platforms your project actually runs on, not all of them.

Supabase MCP — query tables, run SQL, apply migrations, read logs, manage branches. Also available as a hosted OAuth server. If Supabase is your backend, this replaces a lot of dashboard round-trips.
Cloudflare MCP — a family of hosted servers for Workers, KV, R2, D1, DNS. Genuinely useful for infra chores you'd otherwise do in the dashboard.
Vercel MCP (hosted at mcp.vercel.com) — deployments, build logs, env vars, project config. Reading a failed deploy's build log without leaving the terminal is the killer feature.
GitHub MCP — the official server: PRs, issues, code search, actions. Honest take: if you already use the gh CLI, Claude Code can drive that directly and you may not need the MCP layer. It wins when your workflow spans many repos, issues, and reviews.

Reasoning and memory (from the official reference set)

The modelcontextprotocol/servers repo maintains the reference implementations — Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking, Time.

Sequential Thinking — a cheap scratchpad that nudges the model through multi-step problems. Low cost, occasionally a real difference on gnarly refactors.
Memory — knowledge-graph persistent memory. Powerful if you maintain it; in practice most people stop feeding it after week one. Skip unless you'll actually tend it.
Filesystem — skip it in Claude Code. The agent already has native file tools; this server exists for MCP hosts that don't.

Observability and analytics

Sentry MCP — pull real stack traces and issue context into the conversation instead of pasting them. Debugging with the actual error beats debugging with a description of it.
PostHog MCP (hosted at mcp.posthog.com) — query analytics, insights, and feature flags in plain language. Great for "did that ship actually change anything" checks without dashboard archaeology.

Web search

Brave Search MCP — maintained by Brave (the old reference implementation was archived), needs a free API key. Claude Code has built-in web search, so this is mainly for teams that want an independent, controllable search backend.

The restraint rule

Every connected server adds tool definitions to Claude's context and more ways to pick the wrong tool. The practical ceiling is around four to six active servers per project — your deploy platform, Context7, Playwright when you're in the UI, and maybe one observability server. Review the rest with claude mcp list and cut what you haven't used in a week. Security matters too: MCP servers run with real credentials against real systems — we wrote up the risks in MCP security: what enterprise teams need to know.

Not sure whether you even need MCP for a given job, or whether a skill would do? That's a common confusion — see Claude Code plugins vs skills vs MCP, explained.

FAQ

Do MCP servers cost money? The servers themselves are almost all free and open source. Some need an API key to a paid service (Brave Search has a free tier; platform servers use your existing account). The hidden cost is context: more connected tools means more tokens per request.

Where do I find more servers? The official MCP registry is the canonical index, and the reference repo links out to hundreds of community servers. Quality varies wildly — prefer vendor-official servers for anything touching production.

How do I keep my MCP configs organized across machines and projects? That's exactly the scattered-toolkit problem Vibe Coders' Kit exists for — it stores your MCP server configs (with env vars) alongside your agents, skills, and prompts, with 50+ pre-configured servers you can import in one click. See also how to organize your Claude Code setup.