MCP Server Guide 2026: ClawHub, Smithery, and What to Install

TL;DR: Model Context Protocol is the universal connector layer that lets AI assistants reach into your real systems — your database, your codebase, your browser, your Slack. The ecosystem now has 2,000+ community-built servers, organized across four major registries. This post walks through the best servers by category, how to wire them into Claude Code and Claude Desktop, and the personal stack I run across every session.

What Is MCP, and Why Does It Exist?

Before MCP, connecting an AI assistant to an external tool meant one of two things: either the AI vendor hardcoded the integration (slow, limited, proprietary), or you wrote custom tool-calling logic for every API you cared about (fragmented, unmaintainable, per-model).

MCP — Model Context Protocol — solves this at the protocol level. It defines a standard way for AI models to discover, call, and receive results from external tools. The model doesn't need to know how your database works. It just needs to know that a server exists, what tools it exposes, and how to call them. The server handles everything else.

The USB-C analogy holds precisely here. Before USB-C, every device had its own connector. After USB-C, one port connects to monitors, storage, power, audio, and peripherals. MCP does the same for AI: one protocol, thousands of integrations, any model that speaks the spec.

Anthropic published the MCP specification in late 2024. Claude adopted it natively. Within months, the community had shipped hundreds of servers. By early 2026, the ecosystem crossed 2,000 publicly listed servers across all major registries — and the pace hasn't slowed.

The Architecture in 30 Seconds

An MCP setup has three parts:

1. The host — the AI client (Claude Code, Claude Desktop, Cursor, Zed, etc.)
2. The server — a process exposing tools over stdio or HTTP/SSE transport
3. The tool — a discrete function the server exposes (e.g., query_database, create_issue, run_playwright)

The host discovers available tools at session start, then calls them on demand as the model determines they're needed. Results flow back as structured data the model can reason over. The whole loop is transparent — you can watch every tool call happen in real time.

The Major MCP Registries

The registry landscape consolidated quickly. Four platforms now index the majority of production-grade MCP servers.

ClawHub

ClawHub positions itself as the curated tier. Every server listed has passed a basic functionality review, and user ratings surface the servers that actually hold up in production. The install command it generates works directly with claude mcp add. For Claude Code users specifically, ClawHub is where I start any server search.

Smithery

Smithery is the largest raw index. If you're looking for something obscure — a server for a specific SaaS tool, a niche database connector, a custom AI pipeline adapter — Smithery has the highest odds of surfacing it. The one-click install feature generates the right config for Claude Desktop automatically. Docs are pulled from each server's repository and rendered inline, which saves a lot of context-switching.

mcp.run

mcp.run takes a different architectural approach: servers run in a sandboxed cloud environment, so you never install anything locally. You connect once, and the servers are always available, always updated. This trades some customization for significant operational simplicity. For teams where everyone needs the same MCP stack without per-machine setup, mcp.run is compelling.

Glama

Glama focuses on enterprise readiness. Every server in its registry has been reviewed for security posture — no credential leakage, no unexpected network calls, no undocumented side effects. Verified publisher badges mean the code you're running matches the source you reviewed. For production systems where trust matters, Glama is the right starting point.

The Anthropic Official Repository

The official modelcontextprotocol/servers repo on GitHub ships reference implementations for the most common integrations: filesystem, GitHub, Google Drive, PostgreSQL, Slack, Brave Search, and more. These are the servers Anthropic builds and maintains directly. They're conservative by design — fewer features, higher stability, well-documented. Use them as the baseline before reaching for community alternatives.

Best MCP Servers by Category

Two thousand servers is noise. These are the ones that hold up across real workflows.

Developer Tools

Sequential-thinking deserves a specific note. It's not a connector to an external system — it's a reasoning scaffold. It gives the model a tool to think step-by-step through complex problems before acting. The difference in output quality on ambiguous tasks is measurable. I run it in every session.

Playwright is the browser automation standard. Web scraping, UI testing, form filling, screenshot capture — one server, full browser control. Pair it with the filesystem server and you have a complete web research pipeline that writes its own notes.

Databases

The Supabase server stands out because it covers the full platform, not just the database. You can create tables, manage auth policies, deploy edge functions, and inspect storage buckets — all through natural language. For projects built on Supabase, this eliminates most dashboard time.

Cloud and DevOps

The Vercel server is the one I use most. Deployment status, build logs, environment variable management, preview URL inspection — all without leaving the conversation.

Productivity

AI and Memory

Memory servers are the unlock for AI that feels continuous rather than amnesiac. The mem0 server persists facts across sessions — preferences, project context, recurring patterns — so the model starts each session already knowing what matters.

Setting Up MCP in Claude Code

Claude Code manages MCP configuration through the claude mcp command. Servers are stored in ~/.claude.json under projects.<path>.mcpServers.

The Add Command

# Basic server from npm
claude mcp add sequential-thinking -- npx -y @anthropic/mcp-server-sequential-thinking

# Server with environment variables
claude mcp add vercel -e VERCEL_TOKEN=your_token -- npx -y @vercel/mcp-adapter-vercel

# Verify what's installed
claude mcp list

The -- separator is required. Everything after it is the command Claude Code runs to start the server. The -e flag injects environment variables — the correct way to pass credentials without hardcoding them.

My Personal MCP Stack (21 Servers)

I run 21 MCP servers across my Claude Code sessions. The architecture is layered: foundation servers that apply to everything, then project-specific integrations scoped per repository.

Foundation Layer (Global)

Infrastructure Layer (Global)

AI Layer

The total comes to 21 servers, but not all are active in every session. Claude Code's project scoping means servers only load for the projects that need them.

Common Pitfalls and Debugging

The mcp-doctor Tool

The mcp-doctor server exposes diagnostic tools directly to the model. Once installed, you can ask Claude to run a health check on all connected servers. It inspects each server, tests connectivity, validates credentials, and reports status. It's the fastest way to debug a broken stack. Read more in the mcp-doctor breakdown.

Hook Script Failures

If you use Claude Code hooks, run them with explicit bash or node prefixes in settings.json. Claude Code invokes hooks via /bin/sh, which doesn't execute shebangs.

Transport Mismatches

MCP supports two transports: stdio (local process) and HTTP/SSE (remote). Most servers use stdio. If a server requires HTTP/SSE — common for hosted servers on mcp.run — the config syntax differs and needs transport: "sse" plus the URL.

Where the Ecosystem Is Going

Hosted server infrastructure is maturing. mcp.run's sandboxed approach removes local installation entirely.

Enterprise registries with security audits and SLA guarantees are emerging. Glama leads here.

Multi-model MCP is arriving. OpenAI, Google, and Mistral are shipping MCP support. The protocol is becoming the actual standard.

For deeper analysis, the MCP Ecosystem Research page tracks what's shipping each month.

For building your own servers, the MCP Server Architecture Workshop covers the full implementation.

For the protocol-level changes in Claude 2.1, the MCP Revolution post covers what made this ecosystem possible.

FAQ

What's the difference between MCP and function calling?

Function calling is the mechanism inside the model for invoking tools. MCP is the protocol that defines what tools exist, how they're discovered, and how results are returned. MCP uses function calling under the hood — it's the standardized layer on top that makes tools portable across models and clients.

Do I need to run MCP servers locally?

Both local and remote options work. Local servers (stdio transport) run as subprocesses — faster, no latency, but require local installation. Remote servers (HTTP/SSE transport) run on external infrastructure — no setup, but dependent on network. mcp.run specializes in hosted remote servers.

How many MCP servers can I run at once?

Each server is a process consuming 80-200MB of memory. On 12GB of RAM, I keep active sessions to 3-5 servers. Project scoping helps: servers only load for the projects that need them.

Are MCP servers safe?

A malicious server could read and transmit data. This is why registries matter: Glama audits for unexpected network calls and credential handling. The model's tool calls are logged — you can see exactly what each server does.

Can I write my own MCP server?

Yes. The official SDK exists for TypeScript and Python. A minimal server is roughly 50 lines of code. The MCP Server Architecture Workshop covers the full implementation.

Does MCP work with models other than Claude?

In 2026, yes — increasingly. OpenAI's API added MCP support in early 2026. Cursor and Zed support MCP natively. The spec is open and model-agnostic.