Publish from Cursor to a link you can share

Not for running apps. Origin, the product Cursor calls code hosting, is a git forge: it stores and shares code, and its docs cover repositories, pull requests and code browsing rather than serving a site.

The two sit side by side. Keep your code on Origin or GitHub, and publish the built output to host0 when someone needs to open it. Origin is in early beta, and Cursor's docs say it isn't available on the Free or Plus plans.

No. A shared canvas is a read-only snapshot for people on your Cursor team. A host0 link is the app itself, for whoever you choose, and it can hold data people write to.

If your audience is your team and a snapshot is enough, a canvas is the shorter path. Reach for host0 when someone outside the team needs it, or when people should be able to change what's on the page.

host0 serves static files, so publish the build output rather than the source. Tell Cursor to run the build first and point the deploy at the output folder.

• A site or single-page app with a build step: build it, then deploy the output folder, such as dist or out.
• A project that needs a database for lists, boards or forms: ask Cursor to move that data to shared records, which need no server code.
• Real server code, like an Express API or Next.js server routes: host0 can't run it yet. Host that part somewhere else, or ask Cursor whether the page can do without it.

It manages apps that already exist, and it can't deploy. Publishing goes through the skill or the CLI, because uploading files is a job for a command rather than a tool call.

• Look: list apps, inspect one, read its analytics.
• Share: set who can open it, invite and remove people.
• Fix: rename the URL, list versions and restore one, edit or delete records, delete an app.

Cursor asks for approval before it uses an MCP tool by default, so a visibility change shows you its arguments first. The terminal agent reads the same mcp.json as the editor.

{
  "mcpServers": {
    "host0": {
      "url": "https://mcp.host0.ai/mcp"
    }
  }
}

Ask again. The first deploy saves .host0/state.json in that folder, so the next one updates the same app at the same URL and skips the files that didn't change.

The catch is a checkout without that file, like a fresh clone or a new worktree. There the deploy doesn't know your app and creates a new one, so tell the Agent the app's current slug, the part of its URL before .host0.app: "redeploy to k3x9q2m7ta". A display name won't match.

Because a deploy needs the network. Cursor's sandbox blocks network access apart from what your network mode and sandbox.json open, and its built-in defaults are for package managers and language tools.

Approve it when Cursor asks, or allow host0's two hosts once. Uploads go to host0.app and everything else to host0.ai, so add both to allow under networkPolicy in ~/.cursor/sandbox.json for every project, or in the project's own .cursor/sandbox.json. Signing in also saves a key to ~/.host0, outside the workspace, so list that under additionalReadwritePaths or approve that run. We haven't run this combination.

The first publish with no key also prints a link and a short code. Cursor shows them to you, you approve in your own browser, and nothing personal goes into the chat.

Ask Cursor to invite them by email, or add them in the dashboard. They sign in with that address, and removing them cuts access right away.

They don't need Cursor. They do need a host0 account for that email if the app is private; the invitation lets them make one. If you've opened the app to anyone with the link, nobody signs in to read it, and an invite is what lets someone change its shared data.

It should work, but we haven't run it ourselves. A cloud agent runs in its own VM, so it doesn't have the skill or the key from your machine, and you have to give it both.

• The skill: put it in the repo's .cursor/skills/, or turn on Sync Skills for Cloud Agents. Only ~/.cursor/skills/ syncs, so a copy in ~/.agents/skills/ won't reach the VM.
• The key: add HOST0_API_KEY in Cursor's Secrets tab. Cloud agents see secrets as environment variables, and the skill reads that one.
• The app: name the slug in your prompt, since the VM has no .host0/state.json unless it's committed.