Short answer: a remote MCP server is just a URL your AI client talks to over Streamable HTTP, with OAuth instead of an API key. This post walks through a real one, the ChefBear menu server, from install commands for five clients to the tool list and the permission design, so you can see what a small, consumer-facing remote server looks like end to end.
I eat out a lot and photograph menus constantly. Those photos usually die in the camera roll. Connecting a menu app to my assistant turned out to be a nice, concrete way to understand remote MCP servers, because the data is small, personal and easy to reason about.
What is a remote MCP server?
The Model Context Protocol (MCP) lets AI clients such as Claude, ChatGPT, Cursor or Gemini CLI call tools exposed by a server. A remote server runs on someone else's infrastructure: there's nothing to install locally, the transport is Streamable HTTP, and authorization follows the MCP authorization spec, which is OAuth 2.1 with PKCE.
Here's the server used in this walkthrough:
-
URL:
https://chef-bear.com/api/menu-sharing/mcp - Transport: Streamable HTTP
- Auth: OAuth sign-in. No API key, no header to set.
-
Registry name:
com.chef-bear/chefbearin the official MCP Registry. The server card is athttps://chef-bear.com/api/menu-sharing/mcp/server-cardand is listed in/.well-known/ai-catalog.json.
To complete sign-in you need an account in the app and its iPhone app (version 4.4 or later), because that's where the connection is approved.
How do I add a remote MCP server to my client?
Name the server chefbear wherever your client asks.
Claude Code
claude mcp add --transport http --scope user chefbear https://chef-bear.com/api/menu-sharing/mcp
Start a new session, run /mcp, pick chefbear and choose Authenticate. Newer versions also support claude mcp login chefbear.
Codex
codex mcp add chefbear --url https://chef-bear.com/api/menu-sharing/mcp
# if the sign-in doesn't open by itself:
codex mcp login chefbear
Cursor — add to ~/.cursor/mcp.json (all projects) or .cursor/mcp.json (one project):
{
"mcpServers": {
"chefbear": { "url": "https://chef-bear.com/api/menu-sharing/mcp" }
}
}
VS Code
code --add-mcp '{"name":"chefbear","type":"http","url":"https://chef-bear.com/api/menu-sharing/mcp"}'
Gemini CLI
gemini mcp add --scope user --transport http chefbear https://chef-bear.com/api/menu-sharing/mcp
# then, inside Gemini CLI:
/mcp auth chefbear
Claude (web, desktop, mobile) and ChatGPT don't need any of this: the server is in Claude's connector directory and available as a ChatGPT plugin. The server's install page (linked at the end) has one-click links for each client, plus steps for Google Antigravity.
Any other MCP client: add a remote server with the URL above and choose Streamable HTTP and OAuth. One constraint: the client has to finish OAuth with a localhost callback on your own machine. Clients that redirect to a hosted web URL can't connect.
How does the OAuth sign-in work?
When your client starts the OAuth flow, the browser opens a sign-in page that hands off to the phone:
- On an iPhone, tap the button that opens the app. On a computer, scan the QR code with your iPhone camera.
- In the app, review what the client is asking for and tap Allow.
- The client receives its token and you're done.
There is no code to type and nothing to paste into the chat, and the assistant never sees a password. That's a good property in general. "Paste the code here" is exactly the kind of instruction a prompt-injected agent might relay, and the simplest defense is a flow with no code in it.
What tools does the server expose?
| Tool | What it does |
|---|---|
import_menu |
Reads 1–5 menu photos into a private menu (asynchronous) |
get_import_status |
Polls an import until the menu is ready |
save_menu |
Saves a menu the assistant already read from a photo or text |
list_menus |
Lists saved menus, newest first (separate permission) |
get_menu |
Reads one saved menu: dishes, short descriptions, categories, prices |
get_shared_menu |
Reads a menu link someone shared with you |
share_menu |
Turns on a public link for a menu, only when asked |
revoke_menu_share |
Turns that link off; the private menu stays |
connection_status |
Checks the connection and what it's allowed to do |
In Claude (web, desktop and mobile, but not Claude Code) there are three more tools, get_taste_profile, update_taste_profile and recommend_dishes, behind separate taste-profile permissions granted on the approval page.
import_menu + get_import_status is a pattern worth noticing: vision processing of several photos can take longer than a client wants to wait on one tool call, so the server returns a job and the model polls. Models handle this well if the tool description says plainly "call get_import_status until status is ready."
What can I ask once it's connected?
- "Check my menu connection and tell me what you're allowed to do." A good first call, because it hits
connection_status. - "Import this menu photo and tell me which dishes are vegetarian."
- "List my last three menus and summarize each in one line."
- "Read this shared menu link and pick two dishes for someone who doesn't like spicy food."
- "Turn off the share link for last night's menu."
In Claude Code or Cursor this is mostly a curiosity, but it's a handy way to inspect real tool schemas and descriptions before you design your own.
What design choices are worth borrowing for your own server?
If you're building a remote MCP server for a consumer app, a few choices in this one generalize well:
1. Approve in an app the user already trusts. Approval happens in the phone app, not in a web form an agent opened. No shared secret exists for an agent to ask for.
2. Make enumeration a separate scope. Reading a menu you point to (get_menu) and listing everything you've saved (list_menus) are different levels of access. Here, listing is its own grant ("See a list of your saved menus"). The underlying OAuth scopes are separate too (menus:read, menus:list, and so on), which you can see in connection_status.
3. Never make anything public without an explicit ask. share_menu only runs when the user asks for a link, and revoke_menu_share is always available. Revoking the link doesn't delete the private menu.
4. Say what the server won't do. This server never books tables, places orders or takes payments. Writing that into tool descriptions keeps models from overpromising.
5. Publish machine-readable discovery. A server card, an ai-catalog.json entry and an llms.txt that spells out the URL, the auth model and which product this is. Name collisions are real: several unrelated apps share similar names, and an agent asked to install one by name needs a way to pick the right one.
What are the limits?
- The approval step needs an iPhone; there's no Android app.
- Accounts in the mainland China service region can't connect.
- Taste-profile tools are available in Claude only for now.
- Dish details come from the restaurant's menu, so allergens still need to be confirmed with the restaurant.
FAQ
Do remote MCP servers need an API key?
No. The MCP authorization spec is OAuth-based, and many remote servers, this one included, use OAuth only, so there's no key to store.
Why does my client fail with a redirect error?
Some servers only accept OAuth redirects to localhost on your machine. If your client redirects to its own hosted URL, the server will reject it. Check whether the client supports local callbacks or client metadata documents.
What's the difference between Streamable HTTP and stdio?
stdio servers run as a local process launched by your client. Streamable HTTP servers are reached over the network, which is what makes a hosted, OAuth-protected server possible.
Can the assistant share my data without asking?
With this server, sharing is a separate tool call that's only made on request, and any share link can be revoked.
Try it
The install guide with one-click links for every client is on the ChefBear MCP page. If you connect from a client that isn't listed, or the OAuth flow fails somewhere, drop it in the comments. Client quirks around localhost callbacks are exactly what's worth collecting.
Top comments (0)