Browser First Local App

Browser First Local App

Purpose

Default a local tool to a local HTTP service plus a browser UI instead of a packaged desktop shell, unless a native shell is genuinely required. Browser-first removes packaging and code-signing overhead, makes the app inspectable through browser devtools from day one, and keeps debugging in familiar territory (network tab, console, breakpoints). The deliverable is a service with a health endpoint, a launcher that gates the browser-open on readiness, file-based logs, and a clean shutdown that frees the port.

When to use

• A local tool or utility is being built and the shell choice (browser vs native desktop) has not been made — default to browser-first.
• An existing local app needs a launcher that health-gates the browser open and shuts down without orphaned ports or processes.
• A project is moving from a CLI-only interface to a visual local UI and needs a minimal server + browser pattern established.
• A native desktop shell is being proposed for a task that only needs filesystem or local-network access — which a localhost service already covers.

When not to use

• The task is unrelated to runtime / local-app work.
• The work requires production deploys, destructive data actions, or secret disclosure.
• A native shell is genuinely required (see the dedicated section) — then browser-first is the wrong default.
• A narrower skill or an existing project launcher convention already covers the need.

Procedure

1. Decide the shell. Default to local service + browser UI. Choose native only if the task hits a browser limitation listed below. Record the reason either way.
2. Pick the server for the stack (static file server, Express/Fastify, FastAPI/uvicorn, Vite preview) and a port — fixed (e.g. 3000) or auto-selected if conflicts are likely.
3. Add a readiness endpoint (GET /health returning 200 with a tiny JSON body) that reflects real readiness — DB connected, migrations applied — not merely "process is up."
4. Bind to localhost (127.0.0.1) by default. Only bind 0.0.0.0 when another device must reach it, and call that out as an exposure decision.
5. Write a launcher that starts the server, polls /health until ready (with a timeout), opens the default browser cross-platform, tees logs to a file, and traps signals for clean shutdown.
6. Send logs to a stable file path per the project's log policy; keep the console output quiet but useful (one "ready on http://localhost:PORT" line).
7. Restrict CORS on the local API to the app's own origin, and inject the API base URL into the client at startup so an auto-selected port still works.
8. Verify the full cycle: cold start, health passes, browser opens to a working page, Ctrl-C stops the server, and the port is free afterward (no orphan process).

When a native shell IS justified

• Needs OS APIs the browser cannot reach: system tray, global keyboard shortcuts, deep/arbitrary filesystem access, native OS menus, multi-window management.
• Must ship as a signed, auto-updating, fully-offline packaged binary.
• Requires native notifications or protocol handlers beyond what a PWA provides.
• Otherwise prefer browser-first: simpler runtime, easier debugging, no packaging tax.

Concrete checks

• Shell decision (browser-first vs native) is recorded with a one-line rationale.
• A GET /health endpoint exists and returns 200 only when the app is truly ready.
• The server binds localhost by default; any 0.0.0.0 bind is justified.
• The launcher polls health before opening the browser and fails closed (non-zero exit, log pointer) if readiness never arrives.
• Logs land in a file at the policy path; secrets never appear in logs or in the opened URL/query string.
• Shutdown kills the server process and frees the port — verified with lsof -i :PORT returning empty.
• The open command is correct per OS (open / xdg-open / start "").
• The readiness poll has a finite timeout and fails closed with a log pointer.
• The client receives its API base URL at startup rather than hardcoding host/port.
• CORS on the local API is restricted to the app's own origin, not *.
• Each "native shell needed" claim maps to a row in the decision matrix, not a vague preference.

Commands

# Confirm the target port is free before launch
lsof -i :3000 || echo "port 3000 free"

# Cross-platform health-gated launcher
#!/usr/bin/env bash
set -euo pipefail
PORT="${PORT:-3000}"; LOG="${LOG:-./run.log}"
npm run start >"$LOG" 2>&1 & SRV=$!
trap 'kill "$SRV" 2>/dev/null' EXIT INT TERM
for i in $(seq 1 60); do
  curl -fsS "http://localhost:$PORT/health" >/dev/null 2>&1 && break
  sleep 0.5
  [ "$i" = 60 ] && { echo "server not ready, see $LOG"; exit 1; }
done
case "$(uname -s)" in
  Darwin) open "http://localhost:$PORT" ;;
  Linux)  xdg-open "http://localhost:$PORT" ;;
  *)      start "" "http://localhost:$PORT" ;;
esac
wait "$SRV"

# Verify readiness and shutdown manually
curl -fsS -w '\n%{http_code}\n' http://localhost:3000/health
# after Ctrl-C, confirm no orphan holds the port:
lsof -i :3000 || echo "port released cleanly"

// Health endpoint that reflects REAL readiness (Express)
app.get('/health', async (_req, res) => {
  try {
    await db.query('SELECT 1');          // dependency check, not just "process up"
    res.status(200).json({ status: 'ok' });
  } catch (e) {
    res.status(503).json({ status: 'degraded' });
  }
});

# FastAPI equivalent
@app.get("/health")
async def health():
    try:
        await database.execute("SELECT 1")
        return {"status": "ok"}
    except Exception:
        return JSONResponse({"status": "degraded"}, status_code=503)

# Auto-select a free port when a fixed one may be taken (Node one-liner)
PORT=$(node -e 'const s=require("net").createServer();s.listen(0,()=>{console.log(s.address().port);s.close()})')
echo "using port $PORT"

# Confirm the server is bound to localhost, not 0.0.0.0 (exposure check)
lsof -nP -iTCP:"$PORT" -sTCP:LISTEN | rg -q '127.0.0.1|\[::1\]' && echo "localhost only" || echo "WARNING: bound on all interfaces"

# Full cold-start verification in one pass (ready -> open -> shutdown -> port free)
./run.sh & LP=$!
until curl -fsS http://localhost:3000/health >/dev/null 2>&1; do sleep 0.5; done
echo "ready"; curl -fsS -w 'home=%{http_code}\n' -o /dev/null http://localhost:3000/
kill "$LP"; sleep 1
lsof -i :3000 && echo "FAIL: orphan holds the port" || echo "PASS: port released"

# CORS check: the local API should reject a foreign origin
curl -s -o /dev/null -w 'cors=%{http_code}\n' -H "Origin: http://evil.example" \
  -X OPTIONS http://localhost:3000/api/data   # expect a restricted/!200 preflight

# No secrets leaked into the run log
rg -nEi "api[_-]?key|secret|token|password" run.log && echo "WARNING: secret in log" || echo "log clean"

Browser-first vs native (decision matrix)

Common issues & anti-patterns

• Opening the browser before the server is ready. A fixed sleep 2 races startup; on a slow machine the browser loads a connection-refused page. Always poll /health.
• Health check that lies. /health returns 200 the instant the process starts, before the DB pool or migrations are ready, so the first real request fails. Make health reflect dependencies.
• Binding 0.0.0.0 by default. Exposes the dev server to the local network (and anything on it) without intent. Bind localhost unless remote access is a stated requirement.
• Orphaned process on Ctrl-C. The launcher backgrounds the server but does not trap signals, so the port stays held and the next launch fails with EADDRINUSE.
• Secrets in the URL. Passing a token as ?token=... puts it in browser history and server logs. Use headers or a session cookie instead.
• Reaching for Electron reflexively. Wrapping a pure web UI in a desktop shell for a task that needs no OS API — paying the packaging and signing tax for nothing.
• No timeout on the readiness poll. An infinite until curl loop hangs forever if the server crashes on boot. Cap the poll and fail closed with a log pointer.
• CORS opened to * on the local API. A dev convenience that lets any site in the browser call the local service. Restrict the origin to the app's own URL.
• Logging request bodies to the console. Convenient locally, but it spills tokens and PII into a file or terminal scrollback. Log metadata, not payloads.
• Hardcoding http://localhost:3000 in the client. Breaks the moment the port auto-selects. Inject the base URL at startup.

Required output

Return: the shell decision (browser-first vs native) with rationale; the server, port, and health endpoint; the launcher script and its log location; and the cold-start verification result (ready → browser opened → clean shutdown, port freed). Note any native-shell requirement that overrode the browser-first default.

Safety

• Bind to localhost by default; never expose the dev server on 0.0.0.0 without a stated reason.
• Always gate the browser-open on the health check and fail closed if readiness never arrives.
• Ensure shutdown kills the server process and frees the port (no orphans).
• Keep secrets out of logs and out of the URL/query string.

Completion criteria

Done means the app runs as a localhost service with a health-gated launcher, the browser opens only after readiness, logs land at the policy path with no secrets, and shutdown is clean with no leaked port or process — with any native-shell exception documented.