Web Publishing & Deployments

Get a site or app that exists in the sandbox onto a public URL — and decide whether it should go up at all. When to publish is the first half of this skill; how is the second.

Never publish unprompted

Default posture: everything stays a local preview until the user explicitly asks to go live. Previewing is free and re-runnable; publishing is a deliberate, opt-in step that puts code on a URL other people can reach.

Publish only when the user asks in plain terms — "publish this", "make it live", "give me a real/permanent/shareable link", "put this online", "I want to send this to people".

• Never publish unprompted. If the site looks finished and you suspect they'll want it live, offer — "Want me to publish this to a shareable URL?" — and wait for the yes.
• Don't auto-republish after edits. Re-running a local preview is fine any time. Pushing changes to a live URL is not — offer, then wait.
• Don't publish from memory. If you believe a site went live earlier but the current state doesn't confirm it's still up, assume the user took it down on purpose. Re-publish only on a fresh, explicit request.
• Subagents never publish. Approval only routes back to the main thread. Finish the build, verify the preview, hand the project path back.

Taking a published site down

You cannot quietly revoke a live URL, and you must never fake a takedown by overwriting the site with a blank page, a placeholder, or a redirect — that leaves confusing, half-broken state and is not a real unpublish. Point the user at the real control (their deploy platform's dashboard, or stopping the shared sandbox preview), leave the project files untouched, and say plainly that a shared preview URL is public-by-link if they asked for privacy.

Before you publish

1. Verify the live build, not just the source. Start the server, exercise the real URL, and confirm the built output works. Don't grep the source and assume.
2. Run the pre-publish security review — mandatory, every publish. See "Security review" below.
3. Flag runtime-only dependencies that won't survive publishing. API keys set in the dev environment aren't automatically present in a standalone deployment, and anything reaching back into the Neuro OS agent runtime or its connectors has no bridge once deployed. Scan first:

   grep -rn -E "(ANTHROPIC_API_KEY|OPENAI_API_KEY|ELEVENLABS_API_KEY|generate_image|generate_video|generate_audio)" \
     --include='*.py' --include='*.ts' --include='*.tsx' --include='*.js' --include='*.jsx' \
     --exclude-dir=node_modules --exclude-dir=dist . 2>/dev/null

For each hit: refactor it out (bake the output as a static asset at build time — cleanest), provision the credential on the deploy host, or tell the user the feature will be degraded and let them choose.

1. Disclose data-persistence limits. If the app stores user-submitted data in SQLite, an in-memory store, or a local file, that data is not durable on an ephemeral host. Say so plainly before publishing, once per publish, in non-technical terms. For real multi-user persistence use Supabase, with the URL and anon key provisioned as production env vars — never hardcoded into the shipped bundle.
2. Use secure cookies behind the preview proxy. Sites served through the sandbox preview proxy sit behind Neuro OS's domain — set Secure, explicitly-named, properly-scoped session cookies rather than relying on framework defaults.

Security review

Before publishing, run a security-review subagent with the prompt in references/security-review-prompt.md. The checks are mostly grep/bash, so a fast, cheap model is fine. Pass it {{project_path}} (absolute path) and {{context}} (one or two lines on what the site is and whether it handles user data — "public marketing page, no user data" vs "small-team task tracker backed by Supabase") so it can calibrate severity.

• BLOCK (exposed secrets, leaked credentials, critical exploitable vulnerabilities): fix what you can automatically — pull hardcoded keys into env vars, add .env to .gitignore. If a fix needs the user, surface it and stop. Never publish over an unaddressed BLOCK.
• WARN: present it and let the user decide.

---

Once the user has said yes and the checks are clean, there are two worlds, and picking the right one is 90% of the job:

• Instant & throwaway — a live URL in seconds, no account, no login, for a preview/demo the user can click and share. It self-destructs after ~1 hour unless claimed. This is the default when someone just wants to see it live.
• Permanent — a real deployment on the user's own hosting account (custom domain, stays up, redeploys). This needs their account, so it involves them.

Always build the site first (npm run build, etc.) so you're publishing the final output directory (dist/, out/, build/, .next/, …), not source.

Which one?

First, check whether the project already answers this. If the repo targets a host — a Vercel or Cloudflare config, a Dockerfile, a CI deploy step — use that exact workflow. Don't invent a competing one. The table below is for projects with no deploy target of their own.

When in doubt for "just show me it live," reach for the Cloudflare temporary deploy first — it's the only path that needs nothing from the user.

Fast path — instant live URL, no account (Cloudflare)

wrangler deploy --temporary provisions a throwaway Cloudflare account, deploys, and prints a live URL plus a claim URL — all with zero credentials. It is the go-to for handing someone a working link in one turn.

# Build first, then point --assets at the output directory.
# --temporary ONLY works when wrangler is unauthenticated, so clear any creds:
npx wrangler@latest logout 2>/dev/null || true
unset CLOUDFLARE_API_TOKEN CLOUDFLARE_API_KEY

npx wrangler@latest deploy --assets ./dist --temporary --name my-site

Parse the output for the two URLs and hand both to the user:

• the live URL (*.workers.dev) — works immediately, share it,
• the claim URL — valid ~60 min; the user opens it to keep the site by signing into (or creating) a Cloudflare account.

Caveats to state plainly when you hand it over:

• It's temporary. The deployment expires after ~60 min of inactivity; re-running the deploy resets the timer, and unclaimed accounts auto-delete. To make it permanent, the user claims it (or use Vercel instead).
• Requires wrangler ≥ 4.102.0, Node 18+, and a genuinely unauthenticated wrangler (no OAuth login, no CLOUDFLARE_* token env).
• Static assets only on this path (HTML/CSS/JS/images/fonts). Dynamic apps → a Worker (still works with --temporary) or Vercel. Full detail + the browser drag-drop flow: references/cloudflare.md.

Permanent / framework / container (Vercel)

For a site that stays up on the user's account, a framework project that needs a real build, or any backend/container, use Vercel — vercel deploy from the CLI (deployment limits were removed, so it's agent/CI-friendly), the browser Vercel Drop (auto-detects and builds frameworks), or Dockerfile.vercel to run any Dockerfile as an autoscaling function. All of this needs the user's Vercel account/token. Full detail: references/vercel.md.

Handing off — always do this

1. Give the user the live URL first thing (that's the payoff).
2. If it's a temporary deploy, say so in one line + give the claim URL and the ~60-min window. Don't let them assume it's permanent.
3. Offer the permanent path as the natural next step ("want this to stay up on your own domain? I'll set it up on Vercel / you can claim it on Cloudflare").
4. Never publish secrets or private data to a temporary, unauthenticated public URL — anything you deploy this way is world-readable. Treat a public URL as public.

Other providers & going deeper (kortix-marketplace)

This skill covers the two fast paths (Cloudflare + Vercel) inline. For anything beyond that — a different provider (Netlify, Render, Fly.io, AWS, Deno Deploy, GitHub Pages, Railway, …) or a deeper, provider-specific skill for Vercel or Cloudflare (full framework configs, edge functions, DNS/domains, CI) — use the kortix-marketplace skill to search the catalog and import a battle-tested one:

kortix marketplace search "netlify deploy" --json
kortix marketplace search "cloudflare workers" --json
kortix marketplace search "fly.io deploy docker" --json

Reach for kortix-marketplace whenever the user names a provider this skill doesn't cover, or wants richer, dedicated Vercel/Cloudflare tooling than the quickstarts here — import and use the specialized skill rather than hand-rolling it.