How to host an internal tool you built with Cursor
The tool runs on your laptop and your team wants to use it. The four ways to host it, and the three things an internal tool needs beyond a server: sign-in, somewhere to keep data, and knowing who is who.
You opened Cursor on a Tuesday and asked for a tool your ops team has wanted for a year: a place to log vendor issues, assign each one to someone, and see what is still open. By the evening it ran at localhost:3000, and it was good.
Now Dana from procurement wants to use it, and so do four other people. The tool needs a home that isn't your laptop.
This article walks through the ways to host it. We make Spryloom, and it is the last of the four options below, so read that section knowing who wrote it. The other three are real choices, and for some teams they are the better ones.
Running is the easy part
Every host can run a Node app. The question for an internal tool is what else it needs, because a tool for coworkers is not a public website. It holds vendor names, costs and complaints, and everyone who opens it is someone you work with. Beyond a running server, it needs three things.
- Sign-in. A check, before anything is sent, that the person is someone you invited. A link nobody has guessed is not a check: links get forwarded.
- Data that survives. A real database. Files written next to the code tend to vanish on the next deploy.
- Who is who. The tool knows the person marking an issue resolved is Dana, without her typing her name. That is also what lets you say "only admins can delete".
Check each option below against those three.
Option 1: keep it on your laptop, with a tunnel
ngrok, Cloudflare Tunnel and similar tools give localhost an address anyone can reach. One command, and Dana can open it.
For a demo, this is the fastest thing there is. Show the tool in a meeting, let one person try it for an afternoon, then turn it off.
It is the wrong home for a tool people rely on. Close the lid and it is gone for everyone. The address is open unless you set up the tunnel product's own access controls. And if your laptop is the only copy of the data, so is the tool's history.
Option 2: a general host, plus a login
Vercel, Railway and Render all run Node apps well, and Cursor can do much of the setup if you ask. You get somewhere to run it, a stable address and room to grow. If someone on your team is a developer who will look after the tool, this is a good answer.
Of the three needs, a host gives you little out of the box. Vercel can limit a deployment to people with Vercel accounts on your team, but the rest is yours:
- Sign-in. Cursor adds a login: sessions, sign-in emails or passwords, a users table. The most sensitive code in the app, and it must stay correct each time Cursor changes something else.
- Data. Some hosts offer a database you can add. You connect it and keep its password out of the code.
- Who is who. Comes from the login you added, so only as good as that code.
None of this is hard for an engineer. For someone who isn't one, each step is a place to get it slightly wrong without noticing. A login that is slightly wrong is worse than none, because everyone believes it works.
Option 3: a general host, with an identity proxy in front
Instead of a login inside the app, put one in front of it. Cloudflare Access and identity-aware proxies from the large clouds check who someone is before the request reaches the app. The app's code stays free of sign-in, and the proxy can pass the person's email to the app.
This is the safer design, and companies with an IT team often have it already. If yours does, ask them to put the tool behind it. It covers sign-in and who is who. Data is still yours to set up on the host, and the proxy needs your domain, a policy for who is allowed, and someone comfortable configuring both.
Option 4: publish it to Spryloom from Cursor
Spryloom runs apps for teams behind sign-in. It is the identity-proxy design from option 3, with the host and the database included, and set up from inside Cursor.
Add Spryloom once, in Cursor Settings → MCP → Add new MCP server, or in ~/.cursor/mcp.json:
{ "mcpServers": { "spryloom": { "command": "npx", "args": ["-y", "spryloom", "mcp"] } } }
Then ask: "publish this for the ops team, only people I invite". The first time, Cursor asks for your email, Spryloom sends you a link, and you check that the code on the page matches the one Cursor shows. That happens once per machine.
The guidance that makes an app publish on the first try is a short set of rules: read the port from PORT, bind to 0.0.0.0, commit the lockfile, keep data in Postgres. Spryloom writes them down as a skill, and Cursor reads skills, including from the .claude/skills/ folder, so run npx -y spryloom skill in the tool's folder and Cursor picks it up. With the skill, Cursor also writes a spryloom.yaml declaring what the tool needs (without it, the publish tool writes one from what Cursor tells it and what it finds in the folder):
app: name: Vendor Issues description: Log vendor issues, assign them, and see what is still open. runtime: frontend: react backend: node build: npm run build start: node server.js access: visibility: invited data: postgres: true tables: [issues, vendors] uploads: false
- Name and description. Dana's invitation shows both.
- How to build and start it.
- invited: only people you invite get in.
- Its own Postgres, at
DATABASE_URL. The tables are listed on the label page. - No uploads. Keep what matters in the database.
Here is how that covers the three needs.
Sign-in. It happens in front of the app, and the app contains no login code. Invite Dana by asking Cursor, or from a terminal:
spry invite vendor-issues --email dana@acme.com
She gets an email naming the tool and who invited her:
Subject you@acme.com invited you to Vendor Issues
You have been invited to Vendor Issues
you@acme.com invited you to use it.
What it does: Log vendor issues, assign them, and see what is still open.
Open Vendor IssuesIt signs you in as dana@acme.com. It works once, for 7 days; after that, sign in with an emailed link.
There are no passwords. Someone who wasn't invited gets nothing, and you are told that they tried.
Data. The tool gets its own Postgres database through DATABASE_URL, and a coworker can see the declared tables on the label page before opening it.
Who is who. Every request reaches the app with two headers:
const email = request.headers['x-spryloom-email']; // dana@acme.com
const role = request.headers['x-spryloom-role']; // 'admin' or 'user'
Spryloom deletes any copy of these headers that arrives from outside, so the app can trust them. "Resolved by Dana" comes from the header, not a form field.
After that, publishing a change is asking Cursor again. The address stays the same, and so do the data and the people. A bad change goes back with spry rollback vendor-issues.
If the tool doesn't start in Cursor with spawn npx ENOENT, Cursor was opened without your terminal's PATH. Run which npx and put the full path in place of npx in the MCP settings.
What Spryloom doesn't do
- It is for people you know. Everyone who opens the tool signs in. A public website or a product for customers belongs on a general host.
- Sign-in is an emailed link. No passwords, and no Google sign-in.
- No file uploads. The filesystem is read-only apart from a temporary folder. Keep what matters in the database.
- Outside services come from a short list. An app reaches its own database and nothing else, unless it declares a host from the list: Stripe, Slack, GitHub, Linear, Notion, Airtable, HubSpot, Google Sheets and a few AI providers. Anything else, ask us at hello@spryloom.com.
- Roles are admin and user. Finer rules inside the tool are the tool's own code, which Cursor can write using the role header.
- Limits. 10 apps per workspace, and 25 invited people per app.
Which one to pick
For the same question asked from Claude Code, with the options in more detail, see How to share a Claude Code app with your team.
Start here
In Cursor, add the MCP server above, or use any MCP client with npx -y spryloom mcp as the server command. Without an agent, from the tool's folder:
npx -y spryloom publish
Before you send Dana the link, open it in a private browser window, where you are signed in to nothing. You should see a sign-in page, not the tool.