Claude MCP: The Complete Setup Guide & Best Servers for 2026
Connect Claude Desktop and Claude Code to real tools, files, and data — the right way
Claude MCP is Claude's implementation of the Model Context Protocol, Anthropic's open standard for connecting an AI model to the tools, files, and services it needs to actually get work done. Instead of copy-pasting data into a chat window, Claude MCP lets Claude Desktop and Claude Code read your GitHub repo, query your Postgres database, search the web, or file a Linear ticket directly — through a standardized MCP server rather than a custom integration built for every single app. This guide covers exactly how to set up MCP servers in both Claude Desktop and Claude Code, which servers are actually worth installing in 2026, and the security tradeoffs to understand before you connect one.
Key Takeaways
- Claude Desktop and Claude Code both speak MCP, but you configure them differently — a JSON file or one-click bundle for Desktop, the
claude mcpCLI for Code - The official reference repo
modelcontextprotocol/servershas crossed 89,000 GitHub stars, and community directories like PulseMCP now track over 18,000 servers - Vendor-hosted remote servers (Stripe, Notion, Asana) are increasingly replacing local, npm-installed servers — no local process or Node.js required
- Not every MCP server deserves your trust — "tool poisoning" is a real, named attack class, and Claude Code now requires manual approval before running project-scoped servers
THE MCP ECOSYSTEM, AUGUST 2026
Sources: modelcontextprotocol/servers on GitHub, PulseMCP Server Directory
What Is Claude MCP?
The Model Context Protocol (MCP) is an open standard Anthropic released on November 25, 2024, built to solve one specific problem: every AI tool that wanted to read your files, query your database, or call an API needed its own bespoke integration, and none of that work was reusable across models. MCP defines a single client-server interface instead — a Claude MCP server exposes a set of tools, resources, or prompts, and any MCP-compatible client, including Claude Desktop and Claude Code, can connect to it the same way. Anthropic's own engineering team has documented this shift toward standardized, model-agnostic tool access in detail; see their write-up on code execution with MCP.
If you are new to the broader concept of tool-using AI systems, our explainer on what AI agents are covers the fundamentals; if you want a wider directory of servers beyond Claude specifically, see our complete MCP servers directory. For a security-first deep dive, read our dedicated guide to MCP security.
MCP Host
The app you talk to — Claude Desktop or Claude Code — which manages one or more MCP clients.
MCP Client
The connection layer inside the host that speaks the protocol to exactly one server at a time.
MCP Server
The program exposing tools and data — GitHub, a filesystem, a database, or a vendor's hosted API.
Three Ways a Server Reaches Claude
MCP servers connect over one of three transports, and which one a server uses affects how you install it.
STDIO
Runs as a local process on your machine, usually launched with npx or uvx. Best for filesystem and local-tool access.
STREAMABLE HTTP
A remote, vendor-hosted server reached over HTTPS with OAuth 2.1 login. Anthropic's current recommended transport for hosted servers.
SSE (LEGACY)
Server-sent events, an older remote transport still supported for backward compatibility but being phased out in favor of HTTP.
Setting Up MCP Servers in Claude Desktop
Claude Desktop offers three ways to connect a server, from zero-config to fully manual. Start with the first option that covers what you need.
1. Connectors Directory (easiest)
Open the Connectors panel inside Claude Desktop or visit the directory at claude.ai/directory, find a listed connector such as Notion, Slack, or Google Drive, and click Connect. Anthropic reviews these listings and hosts the OAuth flow, so there is no config file to edit at all. See Anthropic's own guide to using Connectors.
2. One-click Desktop Extensions (.mcpb)
For servers not (yet) in the directory, Anthropic ships pre-packaged .mcpb bundles (formerly called .dxt, Desktop Extensions) that install with a single click — no Node.js, no terminal, and secrets are stored in your OS keychain rather than a plaintext file. Read the original announcement on Desktop Extensions.
3. Manual JSON config (most control)
For local servers or anything custom, edit claude_desktop_config.json directly: open Claude Desktop, go to Settings → Developer → Edit Config, then add an entry and restart the app.
~/Library/Application Support/Claude/claude_desktop_config.json%APPDATA%\Claude\claude_desktop_config.json{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-filesystem",
"/Users/you/Documents"
]
}
}
}Use absolute paths, always restart Claude Desktop after saving, and confirm the tool icon (hammer/slider icon) appears in the chat composer. Full reference: Connect local MCP servers on modelcontextprotocol.io.
Setting Up MCP Servers in Claude Code
Claude Code manages MCP servers through the claude mcp CLI rather than a config-file edit, which makes it easier to script and share across a team. Full reference is in Anthropic's Claude Code MCP docs.
Add a remote server (HTTP, recommended)
claude mcp add --transport http notion https://mcp.notion.com/mcp claude mcp add --transport http stripe https://mcp.stripe.com --scope project
Add a local server (stdio)
claude mcp add --env AIRTABLE_API_KEY=your_key \ --transport stdio airtable -- npx -y airtable-mcp-server
Check what is connected
claude mcp list
Status shows as Connected, Needs authentication (click through the OAuth prompt), Failed to connect, or Pending approval for project-scoped servers you have not yet confirmed.
Scopes: who else sees your server
| Scope | Stored In | Visible To | Use Case |
|---|---|---|---|
| local (default) | ~/.claude.json | Only you, this project | Personal API keys, experiments |
| project | .mcp.json (git-committed) | Everyone on the repo | Team-shared tools like GitHub, CI |
| user | ~/.claude.json | Only you, every project | Tools you always want available |
.mcp.json require your explicit interactive approval the first time Claude Code loads them — a deliberate guardrail against a colleague (or an attacker) silently adding a malicious server to the repo.The Best MCP Servers for Claude in 2026
With tens of thousands of listed servers, most people only need a handful. Here are the ones worth installing first, grouped by what they are actually for.
Anthropic Reference Servers
Official, maintained in modelcontextprotocol/servers
- ✓Filesystem — read/write local files and directories you explicitly allow
- ✓Fetch — retrieve and convert web pages for Claude to read
- ✓Git — inspect commits, diffs, and branches in a local repo
- ✓Memory — a simple knowledge graph for persisting facts across sessions
- ✓Sequential Thinking — structured, step-by-step reasoning for complex tasks
Vendor-Hosted (Remote) Servers
Run by the company itself, no local install
- ✓GitHub — issues, PRs, code search, and repo management from Claude Code
- ✓Notion — read and edit pages and databases
- ✓Stripe — inspect payments, customers, and subscriptions
- ✓Sentry — pull error details straight into a debugging session
- ✓Linear / Asana — create and update tickets without leaving the terminal
- ✓Cloudflare — manage Workers, KV, R2, and D1 resources
Popular Community Servers
Widely used, install via npx/uvx
- ✓Playwright / Puppeteer — browser automation and screenshots
- ✓Postgres / SQLite — query a database in natural language
- ✓Brave Search — privacy-respecting web search results
- ✓Context7 — pulls current library documentation into context, reducing outdated API hallucinations
- ✓Figma — read design files and specs for frontend work
For the full, continuously updated list beyond this Claude-specific shortlist, see our MCP servers directory, and if you are chaining multiple MCP servers behind one endpoint, our guide to the MCP gateway pattern covers routing and auth at scale.
Desktop vs. Code: Which Servers Fit Where
Both apps support the same protocol, but the servers people actually attach to each differ by workflow.
| Server | Claude Desktop | Claude Code | Typical User |
|---|---|---|---|
| Filesystem | Common | Rarely needed | Anyone organizing local documents |
| GitHub | Occasional | Core | Engineers reviewing PRs and issues |
| Postgres / SQLite | Occasional | Core | Backend developers debugging data |
| Notion / Google Drive | Core | Rarely needed | Knowledge workers, PMs, writers |
| Playwright / Puppeteer | Occasional | Core | QA and frontend engineers |
| Sentry / Linear | Occasional | Core | Engineers closing the loop from bug to fix |
Is It Safe? MCP Security in Plain Terms
MCP is just a protocol — it does not vet the servers built on top of it. Installing a server means granting it whatever access its tools request, and that access is real: a filesystem server can read files, a database server can run queries, a shell-capable server can execute commands. Security researchers have documented tool poisoning as a named attack class, where a malicious or compromised server hides instructions inside its own tool descriptions, which get pulled into Claude's context the moment the server is listed. See the OWASP MCP Security Cheat Sheet for the full threat model.
Four rules before you connect a server
- Prefer official or vendor-hosted servers (Anthropic, GitHub, Stripe, Sentry) over unfamiliar community packages
- Read what scopes/permissions a server asks for — a "notes" server should not need shell access
- Keep secrets in the OS keychain (via .mcpb bundles or Connectors) rather than pasted into a plaintext config file
- Use project scope in Claude Code deliberately — every teammate will be prompted to approve it, so treat that prompt as a real checkpoint, not noise to click through
For the complete breakdown of MCP threat models, authentication patterns, and enterprise controls, read our full guide to MCP security.
Common Setup Problems
Server does not show up in Claude Desktop
Confirm the JSON is valid (a trailing comma will silently break it), use absolute rather than relative paths, and fully quit and reopen the app — a reload is not enough.
Claude Code shows "Failed to connect"
Run the underlying command manually in your terminal first (for example npx -y some-mcp-server) to confirm it starts without Claude in the loop.
Stuck on "Needs authentication"
Remote HTTP servers use OAuth 2.1 — a browser tab should open automatically. If it does not, check for a blocked pop-up or re-run the add command.
A teammate's project server never loads
Project-scoped servers in .mcp.json require interactive approval per machine — each teammate must confirm it once locally, it does not happen automatically on pull.
MCP connects Claude to your tools. Something still has to run the work.
Wiring up Claude MCP servers gives Claude access to your GitHub, database, or CMS — but someone still has to plan and execute the actual growth work day after day. Planetary Labour is built on the same agentic principle, applied to go-to-market: it autonomously publishes SEO content, posts to X and Reddit, and builds domain authority around the clock, so your GTM keeps moving the way your MCP-connected Claude keeps your codebase moving.
Frequently Asked Questions
What is Claude MCP?
Claude MCP refers to Claude's support for the Model Context Protocol, an open standard Anthropic released in November 2024 that lets Claude Desktop and Claude Code connect to external tools, files, databases, and APIs through a common interface called an MCP server, instead of a one-off custom integration for each service.
How do I add MCP servers to Claude Code?
Run claude mcp add in your terminal — for a remote server, claude mcp add --transport http notion https://mcp.notion.com/mcp; for a local one, add --transport stdio. Check claude mcp list to confirm it connected, and add --scope project to share the config with your team.
How do I add MCP servers to Claude Desktop?
Open Settings → Developer → Edit Config to reach claude_desktop_config.json, add your server entry, save, and restart the app. For most popular tools, skip manual JSON entirely and install a one-click .mcpb bundle or a hosted Connector instead.
Are MCP servers safe to install?
Official and vendor-maintained servers (Anthropic, GitHub, Stripe, Sentry, Notion) are generally safe. Third-party community servers carry real risk — a malicious or compromised server can hide instructions in its tool descriptions, known as tool poisoning — so only install servers from sources you trust and review the permissions they request.
What is the difference between Claude Desktop MCP servers and Claude Code MCP servers?
They use the same underlying protocol and often the same servers, but Claude Desktop targets conversational, everyday use (email, docs, search) configured via JSON or one-click bundles, while Claude Code targets engineering workflows (GitHub, databases, CI) configured via the claude mcp CLI with project, user, and local scopes for teams.