The real difference is the transport
“Local” and “remote” are not official categories. They are shorthand for the two standard transports in the MCP specification:
| Local (stdio) | Remote (Streamable HTTP) | |
|---|---|---|
| Who starts it | Your AI app launches it as a subprocess | It runs on its own, on a server or at a provider |
| How messages travel | Over the process’s standard input and output | Each message is an HTTP POST to one endpoint, such as https://example.com/mcp |
| How many clients | Typically one | Typically many |
| Sign-in | No OAuth; credentials come from the environment | Bearer tokens, API keys or headers; MCP recommends OAuth |
| Where it reaches | Your files, your local network, your installed tools | Whatever the server operator connected |
| Updates | Each person updates their own copy | The operator updates once for everyone |
The MCP documentation gives an example of each: Claude Desktop launching the filesystem server is local, because it runs on the same machine over stdio. The official Sentry server runs on Sentry’s platform over Streamable HTTP, so it is remote.
If you still see “SSE” in older guides: that HTTP+SSE transport from 2024 has been deprecated since protocol version 2025-03-26. Streamable HTTP replaced it, and in the current revision (2026-07-28) it no longer uses protocol-level sessions.
Which AI tools support which
| AI tool | Local (stdio) | Remote (HTTP) |
|---|---|---|
| Claude, claude.ai custom connectors | No | Yes, reachable over the public internet on HTTPS port 443 |
| Claude Desktop | Yes, through its config file | Yes, through your Claude account’s connectors |
| Claude Code | Yes, --transport stdio |
Yes, --transport http (the recommended option for remote servers) |
| Visual Studio Code with Copilot | Yes, “Command (stdio)” | Yes, HTTP |
| ChatGPT | Desktop app: yes, stdio; web: requires a tunnel | Yes; availability depends on the surface and workspace policy |
Note the claude.ai detail: connections come from Anthropic’s cloud, not from your laptop. A server behind a VPN or firewall will not connect unless you allowlist Anthropic’s IP ranges. Ports other than 443 are not supported.
Sign-in works differently
For a remote server, the specification describes a full OAuth 2.1 flow. The MCP server acts as a resource server: it answers a request without a token with HTTP 401 and points to its authorization server. The client opens your browser, you sign in, and the client gets a token issued specifically for that server. A server must reject tokens meant for anything else, so it cannot simply pass your token on to another API.
In practice that looks like this in Claude Code:
- Add the server:
claude mcp add --transport http sentry https://mcp.sentry.dev/mcp. - Run
/mcpinside Claude Code. - Log in through the browser. Claude Code stores the token and refreshes it.
For a local stdio server, the specification says the opposite: do not run the OAuth flow, take credentials from the environment. That usually means an API key in an environment variable in your config, for example claude mcp add --env AIRTABLE_API_KEY=YOUR_KEY --transport stdio airtable -- npx -y airtable-mcp-server. Every colleague then needs their own key on their own machine.
The common misunderstanding: “local is safer”
Local feels safer because nothing leaves your machine. That is only half true. A local server is code you download and run with your own user rights. The MCP security guidance lists what can go wrong: a malicious start command in a config, a malicious package, or an insecure local HTTP server that a website reaches through DNS rebinding. That is why the specification says a server running locally over HTTP should bind only to 127.0.0.1, check the Origin header and require authentication.
A remote server moves the risk to the operator: who runs it, what it logs, and which scopes its tokens carry. Neither is safe by default. For the controls a company needs, read is MCP safe to use in a company.
What this means for a team
- Local servers do not share. A server on one laptop is invisible to the colleague next to you, and to your own ChatGPT in the browser. Each person installs, configures and updates their own copy.
- Remote servers can be shared, with per-person sign-in. Everyone connects to the same URL and logs in as themselves, so the server can apply each person’s rights and log who did what.
- Shared configuration helps, but does not share data. Claude Code’s project scope writes a
.mcp.jsoninto the repository, so the whole team gets the same server list through git. That shares the setup, not what each person’s AI learned. More in Claude Code MCP servers for a team.
A sensible split: local for things that really live on one machine (your files, a local database, a browser for testing), remote for anything the team should see the same way.
Where Bombos fits
Bombos is a plugin that gives every AI tool your team uses (Claude, Claude Code, ChatGPT, Copilot, Grok, Codex, Cursor, your own agents) one shared memory: team knowledge, rules the AI learned about how you work, and what already sits in mail, documents and repos, with a trail of what each answer rests on. It runs in our cloud or on your own server, and one person can try it first with their own agents.
Related questions
Can claude.ai use a local MCP server?
Not as a custom connector. Custom connectors connect from Anthropic's cloud, so the server must be reachable over the public internet on HTTPS port 443. Local servers configured in Claude Desktop's config file are a separate mechanism that works on your own device.
Is the old SSE transport still supported?
The HTTP+SSE transport from 2024 has been deprecated since protocol version 2025-03-26. New servers should use Streamable HTTP; Claude Code also marks SSE as deprecated.
How does a local MCP server get its API key?
The MCP specification says stdio servers should not use the OAuth flow but take credentials from the environment, for example an environment variable set in the client's config.
Can I run an HTTP MCP server on my own laptop?
Yes, but the specification says a locally running HTTP server should bind only to 127.0.0.1, validate the Origin header and use authentication, otherwise websites can reach it through DNS rebinding.
