Most guides make MCP sound like a weekend project. It is not. This guide walks through how to connect MCP server to Claude — on the desktop app, in Claude Code, and through web connectors — and flags the parts worth skipping. By 2026, the Model Context Protocol is the standard way Claude reaches your files, tools, and services, and wiring one up takes a few minutes once you know which path fits.
What changed in 2026
A year ago, almost every MCP setup meant a local process and a hand-edited config file. That still works, but the bigger shift is remote servers. Claude now connects to hosted MCP servers as "connectors," authenticated with OAuth instead of pasted tokens, and the transport moved from the old SSE hookup to Streamable HTTP. There is also a growing directory of one-click connectors, so for popular tools you may never touch a config file. For developers, the API gained a built-in MCP connector so Claude can call a remote server directly from a Messages request. Net effect: less config wrangling, more clicking — but the same trust questions.
Pick the right path
Not every method suits every job. Choose based on where the server runs and who needs it.
| Method |
Best for |
Transport |
Auth |
| Claude Desktop config |
Local tools (files, databases) |
stdio |
Local process |
| Claude.ai connectors |
Hosted SaaS tools |
Streamable HTTP |
OAuth |
| Claude Code |
Coding workflows, per-project |
stdio or HTTP |
Local or OAuth |
| Developer API connector |
Apps you build |
Streamable HTTP |
Token / OAuth |
Connect a local server on Claude Desktop
Local servers run as a process on your machine and talk to Claude over stdio. Open Settings, go to Developer, and edit claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\). Add the server under an mcpServers block:
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/Documents"]
}
}
}
Save, fully quit, and reopen Claude. You should see the server's tools available in the chat. If nothing appears, check that the command runs in your terminal first — most failures are a missing runtime or a bad path.
Connect a remote server with connectors
Hosted servers skip the config file. In Claude's settings, open Connectors, choose Add custom connector, and paste the server's URL. Claude opens an OAuth flow so you approve access with the service's own login — no long-lived token sits in a file. Once approved, the tools show up like any built-in feature. This is the cleanest path for team tools, and revoking access later is a matter of removing the connector rather than deleting secrets.
Wire it into Claude Code
For coding, the CLI is faster than editing JSON by hand. Run claude mcp add <name> -- <command> to register a server, and claude mcp list to confirm it loaded. Claude Code supports scopes: keep a server local to you, share it with a project via a checked-in .mcp.json, or make it available across every project. The project scope is handy for teams, since teammates pick up the same servers when they clone the repo — just make sure the config does not embed anyone's credentials.
What to skip and watch out for
An MCP server sees whatever you connect it to. A filesystem server can read the folder you point it at; a database server can run queries. Treat each one like installing software, because that is what it is.
- Skip unvetted servers. Do not paste a random
command string or token from a gist. Read the source or use an official or verified server.
- Watch for prompt injection. A malicious tool result can try to steer Claude, so keep tool permissions on "ask" for anything that writes or deletes.
- Do not over-scope. Point a filesystem server at one folder, not your entire home directory.
- Ignore the hype counts. Directory listings and "thousands of servers" numbers change constantly — verify what a server actually does before trusting it.
FAQ
Do I need to code to connect an MCP server?
No. Remote connectors are point-and-click, and local servers usually need only a short config block copied from the server's docs.
Why does my server not show up after editing the config?
Almost always a restart or path issue. Fully quit Claude (not just close the window), and confirm the command runs on its own in a terminal.
Is stdio or Streamable HTTP better?
Use stdio for local tools on your own machine and Streamable HTTP for hosted services. SSE is legacy in 2026 — prefer Streamable HTTP for anything remote.
Can Claude connect to more than one server at once?
Yes. You can register several, and Claude picks the right tools per request. Keep the set small so permissions stay easy to reason about.
Where to go next
If you are weighing whether a paid tier is worth it for heavier tool use, read is ChatGPT Plus worth it in 2026. To run models alongside your MCP tools without a subscription, see our local LLM setup guide for 2026. And if you are calling Claude through the API with MCP connectors, our guide on how to reduce AI API costs in 2026 will keep the bill sane.