servelink-swyftlabs/serve-mcp

**You already ask your AI to make things. Now ask it to share them.**
That's the product. You say a sentence, and the thing you just made has a
real URL — one you can text to a friend, drop into a Slack thread, or put on
a slide. It works the same whether you built a Next.js app or asked for a
sales deck, because serve doesn't care what the thing is. It only cares that
you want someone else to see it.
**serve is a hosted service.** Create an account, log in once, and your links
live at *.servelink.cc.
> **About this repository.** This is the public listing and configuration home
> for serve: the MCP registry manifest, the Claude Code plugin manifest, and the
> user documentation. serve is a commercial hosted service and its
> implementation repository is private. Everything you need to *use* serve is
> here or on npm — there is no build step, and nothing here needs compiling.
---
A link can live in two places
Every serve link is either **live from your machine** or **kept alive on
ours**.
- **Live from your machine.** The link points at whatever is running on your
computer right now. Change a heading and the person holding the link sees
the new heading. Close your laptop and the link goes quiet. (Under the hood
this is a tunnel; your agent opens it for you.)
- **Kept alive on ours.** We hold a copy on serve's servers. Your laptop can
be shut, asleep, or in a bag at the airport, and the link still opens.
You don't choose up front, and you're never stuck with the choice. "Serve
this so my teammate can watch while I work on it" gives you the first one.
"Keep it up so they can read it tomorrow" gives you the second. Same link,
same URL, one sentence apart.
That switch is the part people don't expect. Everywhere else, "let someone
see what's on my laptop" and "put this on the internet" are two different
products with two different setups, and moving between them is a project.
Here they're two ways of phrasing the same request.
You never operate it — your agent does
serve's real interface isn't the terminal. It's MCP, the protocol AI agents
already speak, which means making a link isn't a task you go and do. It's
something you mention. Your agent creates the link, tells you the URL, puts
a password on it if you ask, and takes it down when you're finished. There's
a CLI too, and it's a good one, but it's there for the times you'd rather
drive manually.
This is also why serve isn't only for developers. If you can get Claude to
build you a dashboard, you can get Claude to hand you the link to it. The
skill it takes to share the thing is the one you already used to make it.
Quick start
Then tell your agent it exists. Anything that speaks MCP will do — Claude Code,
Claude Desktop, Cursor, opencode:
That's the whole setup. From here it's conversation: ask for a link, get a link.
If you installed globally with npm i -g, "command": "serve" with no args
works too and starts marginally faster.
Prefer Homebrew?
One account works across as many machines as you like.
Claude Code plugin
This repository is also a Claude Code plugin marketplace, so you can skip the
JSON:
Or point Claude Code at a local clone while you try it:
The plugin ships the MCP server configuration in .mcp.json and
nothing else — no hooks, no agents, no code that runs on your machine beyond the
serve client itself.
Things you can just ask for
None of these are commands. They're the kind of thing people actually say, and
the agent works out the rest.
> "Serve me this so I can check it on my phone."
> "Put the deck somewhere my client can open it, and password-protect it."
> "This link is going in an email, so make sure it still works next week."
> "Take down everything I've got open."
> "What am I sharing right now?"