The MCP Servers Worth Adding to Claude Code (2026)

There are thousands of MCP servers now, and the standard advice — "install more servers, unlock more power" — is exactly backwards. Every server you add registers its tool descriptions into your agent's context on every session. A bloated MCP setup makes Claude Code dumber per token before you type your first prompt.

So this is not a list of 25 servers. It's the short setup that earns its context cost in daily use, the rule for evaluating everything else, and the one capability gap no ordinary MCP server can close.

Rule zero: don't duplicate the native toolset

Claude Code already reads files, edits files, greps, runs shell commands, and drives git. Filesystem servers, shell servers, and basic git servers — the most-installed category on every directory — mostly re-sell what's already built in, at the price of extra context and an extra process. Skip them unless you have a specific sandboxing reason.

The bar for adding a server: it must do something the agent genuinely cannot do natively. Four servers clear that bar for most developers, and they're all first-party.

The core four

Playwright — a real browser. Claude Code can write frontend code but can't see it. Microsoft's Playwright MCP lets the agent drive a browser, click through flows, read the console, and verify its own work instead of declaring victory blind:

claude mcp add playwright -- npx @playwright/mcp@latest

GitHub — the parts git can't reach. The git CLI already covers commits and branches. GitHub's official MCP server adds what lives on the platform instead: issues, PR reviews, Actions runs, cross-repo search. It runs as a remote server:

claude mcp add --transport http github https://api.githubcopilot.com/mcp/ \
  -H "Authorization: Bearer YOUR_GITHUB_PAT"

(OAuth sign-in also works in some clients, but token auth is the reliable path in Claude Code today.)

Context7 — docs from this year, not training data. The model's knowledge of your framework froze at its training cutoff; your dependencies didn't. Context7 pulls current, version-specific library docs into context on demand, which kills a whole class of "that API changed" hallucinations:

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

Memory — continuity between sessions. The reference memory server gives the agent a persistent knowledge graph so project decisions survive the end of a session:

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

(Newer Claude Code versions also ship native per-project memory — if yours does, apply rule zero and skip this one.)

The gap none of them close: your agent can't pay for anything

Look at what the core four have in common: they extend the agent's hands — but every external capability behind them is either free or billed to an account you set up. The moment your agent needs something priced — a second opinion from a non-Claude model, an image render, a video, real-time search, market data — you're back to the pre-agent workflow: create an account per provider, generate an API key per account, paste secrets into config, repeat.

That's the actual bottleneck of agent capability in 2026, and it isn't a tooling problem — it's a payments problem. Agents can't fill signup forms and can't hold credit cards. What they can do is sign transactions.

That's the gap BlockRun MCP closes. One server, one wallet, zero API keys:

claude mcp add blockrun -s user -- npx -y @blockrun/mcp@latest

From inside Claude Code, the agent can then call nearly a hundred models (GPT, Gemini, Grok, DeepSeek — useful precisely when you want a non-Claude opinion), generate images, video, and music, run web and X search, query prediction markets and crypto data, and hit multi-chain RPC — each call paid individually in USDC via x402 from a local wallet, priced per request before you commit. No provider accounts, no key rotation, no subscription idling at $20/month.

Honesty section, because this is our product: calls cost real money (per-request, shown up front; several chat models are entirely free, no wallet needed), and giving an agent a funded wallet means deciding how much it may spend — start with a few dollars, not your treasury.

How to evaluate everything else

For any other server a directory recommends, ask three questions:

1. Does it beat the native tool? (Rule zero. Most don't.)
2. What's the context cost? A server registering fifty tools you'll use twice is a tax on every request. Prefer servers that do one thing, or that let you scope which tools load.
3. Do you trust it with your conversation? An MCP server sees the tool calls the agent makes and can shape what the agent believes. Treat adding a server like adding a dependency with network access — because that's what it is. Stick to first-party servers (Microsoft, GitHub, Anthropic's reference set) and vendors you'd trust with an API key.

For domain-specific work, curated lists beat directories: for finance and crypto agents we maintain awesome-finance-mcp, a reviewed list of market-data, trading, and on-chain servers.

FAQ

How many MCP servers should I run? Fewer than you think. The setups that survive daily use tend to hold four to six servers with clear jobs. If you can't name the last time a server's tools were called, remove it.

Do MCP servers slow Claude Code down? Local npx servers add little runtime latency, but every registered tool description consumes context on each request. The cost is thinking room, not milliseconds.

Can an MCP server spend my money? Only one designed to — and that's a feature when it's deliberate. BlockRun MCP pays from a local wallet you fund and cap; every call settles on-chain with a receipt, so an agent's spending is fully auditable. That's a stronger control than a shared API key with monthly billing.

What about writing my own? If your agent needs your internal systems, a small custom server is often better than three generic ones. The MCP spec is deliberately small; a useful server is an afternoon of work.