This lesson is part of Learn MCP, our free course on the Model Context Protocol, within The MCP Guide.
In this lesson you will:
- Compare remote and local MCP servers on setup, credentials, security and team use.
- Learn which AI apps can use each kind.
- Practice choosing remote or local for four real-world scenarios.
Lesson 5 explained the two transports in theory. This lesson is about the choice you face in practice. The same tool is often offered both ways, and the right pick depends on who uses it, what it touches and which app you run.
Key Takeaways
- A local server runs on your computer, usually over stdio, and typically serves one client.
- A remote server runs on the internet over Streamable HTTP and can serve many clients.
- Local servers run with your user permissions, so they can reach anything you can.
- Remote servers usually use OAuth sign-in; local servers usually read keys from environment variables.
- Claude's web app and Cowork can't use local servers from the desktop config file.
- For most business teams, an official remote server is the simpler default.
1. The difference in one table
The MCP architecture overview says local servers that use stdio "typically serve a single MCP client." Remote servers that use Streamable HTTP "will typically serve many MCP clients." Here is how that plays out:
| Factor | Local server | Remote server |
|---|---|---|
| Where it runs | On your computer, started by your AI app | On the vendor's or your company's infrastructure |
| Transport | stdio (standard input/output) | Streamable HTTP |
| Setup | Install a runtime, edit a config file or run a command | Paste a URL and sign in |
| Credentials | Usually API keys in environment variables | Usually OAuth sign-in per user |
| What it can reach | Files and programs your user account can reach | What the signed-in account can reach |
| Works offline or on private files | Yes | Only if the server can reach them |
| Team rollout | Each device needs its own setup | One URL for everyone |
| Updates | You update the package | The provider updates the server |
See MCP transports explained for the protocol details.
2. Setup and maintenance
The remote servers guide gives the main advantage of remote servers. Unlike local servers that need installation on each device, remote servers are available from any MCP client with an internet connection.
Local servers take more work. The local servers guide asks you to install Node.js, edit a JSON config file and restart Claude Desktop. That is fine for one developer. It is slow across fifty laptops.
3. Credentials and sign-in
Remote servers usually use OAuth, the "sign in with" style flow. In Claude Code, the MCP docs say to use /mcp to authenticate with remote servers that require OAuth, then log in through your browser. Cursor's docs say its sign-in uses the authorization code flow with PKCE. Our guide to MCP authentication and OAuth explains the flow.
Local servers usually take an API key in an environment variable. Here is the Claude Code example for a local Airtable server, from its docs:
claude mcp add --env AIRTABLE_API_KEY=YOUR_KEY --transport stdio airtable \
-- npx -y airtable-mcp-serverCursor's docs advise keeping secrets in environment variables, not hardcoded, and using keys with minimal permissions.
4. Security trade-offs
Local servers run as you. The Claude Desktop guide warns that a local server "runs with your user account permissions, so it can perform any file operations you can perform manually." Only give it the folders it needs. Sandboxing local MCP servers covers ways to limit it further.
Local servers should not listen on the network. The spec's Security Best Practices page says servers meant to run locally SHOULD use stdio to limit access to just the MCP client. If a local server uses HTTP instead, the Streamable HTTP spec says it SHOULD bind only to localhost (127.0.0.1). Servers MUST also validate the Origin header to prevent DNS rebinding attacks. Read exposed local MCP servers.
Remote servers see your data in transit. Your requests and results pass through the provider's servers. Review the provider's privacy terms, the scopes it requests, and who operates it. A remote server on an unknown domain deserves the full Lesson 9 review.
5. Which apps support which
Support varies, as of October 2026. Anthropic's custom connector guide says local servers from claude_desktop_config.json are a separate mechanism. They aren't available in Cowork or claude.ai. The same guide says remote servers must be reachable over the public internet from Anthropic's IP ranges.
Coding tools such as Claude Code, Cursor and VS Code support both local and remote servers, according to their docs. For a full rundown, see AI apps that support MCP.
6. A simple decision rule
Use this table as a starting point:
| Situation | Usually choose | Why |
|---|---|---|
| A whole team needs the same SaaS tool | Remote | One URL, individual sign-in, central updates |
| You use Claude on the web or mobile | Remote | Local desktop config servers aren't available there |
| You need files on your own laptop | Local | The server must run where the files are |
| The data must not leave your network | Local, or a remote server you host privately | Keeps traffic inside your boundary |
| You are testing or building a server | Local | Fast to change and restart |
| The vendor only ships one option | That option | Don't build a workaround without a reason |
When a vendor offers an official remote server, it is usually the simpler choice for business users. When you need local files or private systems, a local server or an internal remote server is the answer. Our MCP development team builds both kinds.
Try it
Decide local or remote for each scenario. Write down your reason, then check the suggested answers.
- A marketing team of eight wants Claude to read their project management tool, which offers an official remote MCP server.
- A developer wants Claude Code to search and edit files in a project on their laptop.
- An accountant uses Claude in a web browser and wants it to read files from a folder on their office PC.
- An operations lead wants an AI app to query an internal database that is only reachable on the company VPN.
Suggested answers:
- Remote. One URL, each person signs in as themselves, and nothing to install on eight devices.
- Local. The files are on the laptop, and a stdio server limits access to Claude Code.
- Neither, as described. Claude on the web can't use local desktop config servers. Use Claude Desktop with a local server, or move the files to a service that has a remote server.
- A remote server hosted inside the company network, or a local one. For Claude connectors, Anthropic's guide says private or VPN-only servers need Anthropic's IP ranges allowlisted. Talk to IT before choosing.
Recap
Local servers run on your machine over stdio, reach what you can reach, and need setup on each device. Remote servers run on the internet over Streamable HTTP, use per-user sign-in, and suit teams. Keep local servers off the network, keep keys in environment variables, and review remote providers like any vendor. This lesson closes the Using MCP module. The next module, starting with setting up your development environment, shows how to build your own server. You can return to the full course at Learn MCP.
Course navigation
Previous lesson: Lesson 11: Rolling MCP Out to a Team.
Next lesson: Lesson 13: Set Up Your Development Environment.
Back to the Learn MCP course outline.
Frequently asked questions
Is a remote MCP server less safe than a local one?
Not by default. A local server runs with your user permissions on your computer, which is a real risk. A remote server sends your data through the provider. Each needs a review; the risks are just different.
Can I use a local MCP server with Claude on the web?
Not through the desktop config file. Anthropic's guide says local servers from claude_desktop_config.json aren't available in claude.ai or Cowork. Use Claude Desktop or a coding tool such as Claude Code for local servers.
Why do some local servers use HTTP instead of stdio?
Some tools are easier to run as a small local web service. If you run one, the spec says it should bind only to localhost (127.0.0.1) and must validate the Origin header. The spec's security page recommends stdio for servers meant to run locally.
Do remote MCP servers need OAuth?
Not always, but it is common. Claude's custom connector settings also allow fixed request headers, such as an API key, when a server does not use OAuth. Per-user OAuth sign-in is usually better for teams because each person keeps their own access.
Can I host a remote MCP server privately for my company?
Yes, but your AI app must be able to reach it. For Claude connectors, the server must be reachable from Anthropic's IP ranges, so private servers need those ranges allowlisted. Coding tools running on your own network may reach internal servers directly.
