How do I add an MCP server to Claude Code?

Two main shapes. Remote/HTTP: claude mcp add --transport http (then authenticate with /mcp if it needs OAuth). Local/stdio: claude mcp add -- npx -y [args], where everything after the -- is the command that runs the server. Manage them with claude mcp list, claude mcp get , and claude mcp remove , or the /mcp panel in a session.

How many MCP servers should I run at once?

Start with 4–8. Claude Code's Tool Search (lazy loading) is on by default, so tool definitions don't all load into context up front — but every connected server still adds some overhead and surface area. Connect what you actually use, scope project-shared ones in .mcp.json, and remove servers you've stopped using.

Are MCP servers safe? What about permissions and prompt injection?

MCP servers can read and act on real systems, so treat them like granting access. Prefer read-only configurations where possible, only add servers you trust, keep tokens in environment variables (not in the command or chat), and review project-scoped .mcp.json servers before approving them — Claude Code prompts for approval on those by design.

Can I use the same MCP server across Claude Code, Cursor, and other tools?

Yes — MCP is an open, client-agnostic standard, so the same server works across Claude Code, Cursor, and other MCP-capable clients. The server is the same; only the way you register it (the CLI command or config file) differs per client.

Best MCP Servers for Claude Code in 2026 (+ Exact Setup Commands)

Skills make Claude Code consistent. MCP servers make it capable — they connect it to GitHub, your database, a browser, live docs, your error tracker. Connect a server whenever you catch yourself copying data into chat from another tool.

This is the servers-and-setup guide. For how skills and plugins work, see the skills & plugins guide; for project structure, the Claude Code projects hub.

Every claude mcp command below was verified against the official Claude Code MCP docs on 22 June 2026. Where a server's install varies (remote vs local, OAuth), I link its repo for the current exact line rather than risk a stale one.

What MCP is (in one paragraph)

The Model Context Protocol is an open standard for connecting AI tools to external systems. An MCP server exposes a system (GitHub, Postgres, a browser) as tools Claude can call. It's client-agnostic — the same server works in Claude Code, Cursor, and other MCP clients. Skills are knowledge; MCP servers are reach.

The MCP servers worth connecting

Curated and checked against live GitHub on 22 June 2026.

Server	What it does	Repo (stars)
Filesystem	Sandboxed read/write outside the working dir	`modelcontextprotocol/servers` (~88k)
GitHub	PRs, issues, code search, repo actions	`github/github-mcp-server` (~31k)
Context7	Live, version-specific library docs (less hallucination)	`upstash/context7` (~58k)
Playwright	Real browser automation + E2E testing	`microsoft/playwright-mcp` (~34k)
Supabase / Postgres	Query tables, inspect schema, run migrations	`supabase/mcp` (~3k)
Sequential Thinking	Structured multi-step reasoning	`modelcontextprotocol/servers`
Memory	Persistent knowledge graph across sessions	`modelcontextprotocol/servers`
Fetch	Lightweight web content fetching	`modelcontextprotocol/servers`
Sentry	Pull error stack traces into a session	`getsentry/sentry-mcp`

The honest starter four: Filesystem + GitHub + Context7 + Playwright covers most real coding work — edit files, manage your repo, get accurate library docs, and verify in a real browser. Add Supabase/Postgres if you have a database.

Exact setup — the verified commands

There are two shapes: remote (HTTP) and local (stdio).

Remote / HTTP server (then authenticate with /mcp if it uses OAuth):

claude mcp add --transport http <name> <url>
# real example from the docs:
claude mcp add --transport http stripe https://mcp.stripe.com

Local / stdio server — everything after -- is the command that runs the server:

# Context7 — live library docs
claude mcp add context7 -- npx -y @upstash/context7-mcp

# Filesystem — give it the path it may touch
claude mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem ~/projects

# Playwright — browser automation
claude mcp add playwright -- npx @playwright/mcp@latest

# Sequential thinking
claude mcp add sequential-thinking -- npx -y @modelcontextprotocol/server-sequential-thinking

Passing secrets — use --env, and keep at least one option between --env and the name (the CLI quirk):

claude mcp add --env SOME_TOKEN=xxxx --transport stdio myserver -- npx -y some-mcp-server

Complex config — add-json (the type accepts streamable-http as an alias for http):

claude mcp add-json context7 '{"type":"stdio","command":"npx","args":["-y","@upstash/context7-mcp"]}'

Manage them:

claude mcp list            # all servers + status
claude mcp get context7    # details (or ⏸ Pending approval / ✗ Rejected)
claude mcp remove context7

In a session, /mcp shows tool counts per server, lets you authenticate OAuth servers, and reconnects dropped ones (HTTP/SSE auto-reconnect with backoff; stdio doesn't).

For GitHub and Supabase specifically, the current install (remote, OAuth, or local) is best taken from the source — see github/github-mcp-server and supabase/mcp. The claude mcp add shape is the same; only the URL/package differs.

Scopes — where the config lives

--scope local (default) — just you, this project.
--scope project — writes to a version-controlled .mcp.json so your team gets it. Claude Code prompts each user for approval before using project servers (a security feature; reset with claude mcp reset-project-choices).
--scope user — available across all your projects.

claude mcp add --transport http stripe --scope project https://mcp.stripe.com

Keeping many servers without context bloat

Tool Search is on by default — Claude Code lazy-loads tool definitions instead of dumping them all into context, so a few servers cost far less than they used to. (Disable only if you must, via ENABLE_TOOL_SEARCH=false.)
Start with 4–8, not 20. Connect what you use.
Prefer read-only configs where a server supports it.
Keep tokens in --env, never in chat.
Scope deliberately — project servers for the team, user scope for your always-on tools.

Put it to use

Wire MCP into a properly structured project: the Claude Code projects hub.
Pair it with the right skills: the skills & plugins guide.
New to the tool? The beginner tutorial.
Built something with a slick MCP stack? List it free on SaaSCity.

FAQ

How do I add an MCP server? Remote: claude mcp add --transport http <name> <url>. Local: claude mcp add <name> -- npx -y <package>. Manage with claude mcp list/get/remove or /mcp.

How many should I run? 4–8 to start. Tool Search (default on) keeps the context cost low, but connect only what you use.

Are they safe? They access real systems — prefer read-only, trust the source, keep tokens in --env, and review project-scoped servers before approving.

Cross-client? Yes — MCP is an open standard; the same server works in Claude Code, Cursor, and other MCP clients.