feat(ai-proxy): run the AI proxy on the CVM host, ready to switch over

Same proxy as netlify/functions/ai.mjs, rebuilt as a standalone Bun service on the CVM host so AI calls stop spending Netlify compute/bandwidth credits and the sats-paid Routstr key lives on our box. - ai-proxy/server.ts: Bun.serve on 127.0.0.1:8787 — NIP-98 + Pro gate, model allowlist, per-IP throttle, /api/ai + /api/ai/status. Own tsconfig (typecheck:ai in npm test); eslint-ignored like cvm-server. - ai-proxy/{package.json,.env.example,skatespots-ai.service,Caddyfile,DEPLOY.md}. - scripts/deploy-ai-proxy.sh (npm run deploy:ai): scp + bun install + unit + restart, mirroring deploy-cvm.sh. - Client: one place decides the AI base, src/lib/sasquatchAi.ts (also the NIP-98 u tag). Defaults unchanged (/api/ai on Netlify); set VITE_SASQUATCH_AI_BASE to move it. Fully reversible. Deployed and verified on the host: /api/ai/status returns the live Routstr balance. Public TLS (DNS + Caddy) is the remaining step before switching the app over, so no client behaviour changes yet.
2026-09-28 13:32:05 +07:00 · 2026-09-28 13:32:05 +07:00 · ec06a63b01
commit ec06a63b01
parent f46f6a8967
15 changed files with 566 additions and 18 deletions
--- a/AGENTS.md
+++ b/AGENTS.md
@ -557,6 +557,23 @@ session with **no other memory of this project**. So:
  host costs a moment, not a blank spot. Blossom is content-addressed, so the
  mirror URL is derived from the sha256 in the path (handles `/<sha>` and
  `/<sha>.<ext>`).
+- **AI proxy on our own host (`ai-proxy/`, 2026-09-28):** the same proxy the
+  Netlify Function runs, rebuilt as a standalone **Bun** service on the CVM host so
+  AI calls stop spending Netlify compute/bandwidth credits and the Routstr key
+  lives on our box. `ai-proxy/server.ts` (`Bun.serve` on 127.0.0.1:8787; NIP-98 +
+  Pro gate + model allowlist + per-IP throttle), deployed with **`npm run deploy:ai`**
+  (`scripts/deploy-ai-proxy.sh`) to `/home/ubuntu/skatespots-ai`, systemd unit
+  `skatespots-ai`, type-checked by `ai-proxy/tsconfig.json` (`typecheck:ai`, wired
+  into `npm test`); `ai-proxy/` is eslint-ignored like `cvm-server/`. Env
+  (`SASQUATCH_*`) copied from Netlify env to `~/skatespots-ai/.env` (600);
+  **verified live** (`POST /api/ai/status` → the real Routstr balance). The client
+  picks the base in **one place — `src/lib/sasquatchAi.ts`** (`sasquatchAiBase()`,
+  also used for the NIP-98 `u` tag): set
+  `VITE_SASQUATCH_AI_BASE=https://ai.skate-spots.org/api/ai` to switch the app
+  over; unset keeps the Netlify Function (fully reversible). ⚠️ **Not switched
+  over yet** — still needs a DNS name → the CVM host + **Caddy** for TLS
+  (`ai-proxy/Caddyfile`, `ai-proxy/DEPLOY.md`); until then the app calls `/api/ai`
+  on Netlify.
 - **Mints directory**: `/mints` page lists Cashu mints from the community
  **MintRadar** project. MintRadar has **no CORS**, so it's fetched same-origin
  at `/api/mints/*` — prod proxies via `public/_redirects` (`/api/mints/* →
--- a/ai-proxy/.env.example
+++ b/ai-proxy/.env.example
@ -0,0 +1,14 @@
+# SkateSpots AI proxy — environment (copy to .env on the CVM host; never commit).
+#
+# The upstream is OpenAI-compatible. For sats-paid inference point it at a Routstr
+# node (NOT the aggregator api.routstr.com, which is down).
+SASQUATCH_AI_BASE=https://routstr.githappens.space/v1
+SASQUATCH_AI_KEY=
+# Comma-separated model allowlist — must be models the node actually serves.
+SASQUATCH_MODELS=
+# 1 = require a NIP-98 identity plus an active Pro pass (platform owners bypass).
+SASQUATCH_REQUIRE_PRO=1
+
+# Bind to localhost only; Caddy terminates TLS and reverse-proxies here.
+AI_PROXY_HOST=127.0.0.1
+AI_PROXY_PORT=8787
--- a/ai-proxy/Caddyfile
+++ b/ai-proxy/Caddyfile
@ -0,0 +1,18 @@
+# Caddyfile for the AI proxy host.
+#
+# Replace the site address with the DNS name you point at this server (for
+# example ai.skate-spots.org). Caddy obtains a Let's Encrypt certificate
+# automatically via the HTTP-01 challenge, so ports 80 and 443 must be reachable
+# from the internet (open them in the provider firewall too, not just ufw).
+#
+# Install + run:
+#   sudo apt install -y debian-keyring debian-archive-keyring apt-transport-https curl
+#   curl -1sLf 'https://dl.cloudsmith.io/public/caddy/stable/gpg.key' | sudo gpg --dearmor -o /usr/share/keyrings/caddy-stable-archive-keyring.gpg
+#   curl -1sLf 'https://dl.cloudsmith.io/public/caddy/stable/debian.deb.txt' | sudo tee /etc/apt/sources.list.d/caddy-stable.list
+#   sudo apt update && sudo apt install -y caddy
+#   sudo cp Caddyfile /etc/caddy/Caddyfile && sudo systemctl reload caddy
+
+ai.skate-spots.org {
+	encode gzip
+	reverse_proxy 127.0.0.1:8787
+}
--- a/ai-proxy/DEPLOY.md
+++ b/ai-proxy/DEPLOY.md
@ -0,0 +1,73 @@
+# SkateSpots AI proxy — placement on the CVM host
+
+Why: the Netlify Function (`netlify/functions/ai.mjs`) works, but every AI call
+spends Netlify **compute + response bandwidth** credits. This service is the same
+proxy running on the CVM host we already own, so AI traffic never touches Netlify
+and the Routstr key lives on our box.
+
+```
+browser / APK  ──HTTPS──▶  Caddy (TLS)  ──http──▶  127.0.0.1:8787  (Bun)
+                                                        │
+                                                        └──HTTPS──▶  Routstr node (sats)
+```
+
+## What's needed (once)
+
+1. **A hostname** pointing at the CVM host, e.g. `ai.skate-spots.org` → the
+   host's public IPv4. Add it wherever `skate-spots.org` DNS is managed.
+2. **Ports 80 + 443 open** to the internet (firewall at the provider; `ufw` is
+   inactive on the host).
+3. **Caddy** for automatic TLS (see `Caddyfile` for the install snippet).
+
+## Deploy
+
+```bash
+npm run deploy:ai          # uploads server.ts + package.json + unit, installs, restarts
+bash scripts/deploy-ai-proxy.sh --logs
+```
+
+The **first** deploy needs `/home/ubuntu/skatespots-ai/.env` to exist (the unit
+reads it). Create it from `ai-proxy/.env.example` with the real Routstr key — that
+key is the one currently in Netlify env `SASQUATCH_AI_KEY`:
+
+```bash
+ssh <host> 'cat > ~/skatespots-ai/.env' <<'EOF'
+SASQUATCH_AI_BASE=https://routstr.githappens.space/v1
+SASQUATCH_AI_KEY=<the sk-… Routstr session key>
+SASQUATCH_MODELS=
+SASQUATCH_REQUIRE_PRO=1
+AI_PROXY_HOST=127.0.0.1
+AI_PROXY_PORT=8787
+EOF
+ssh <host> 'chmod 600 ~/skatespots-ai/.env'
+```
+
+## Point the app at it
+
+The client resolves its AI base in one place (`src/lib/sasquatchAi.ts`). Set the
+build-time override and redeploy the web app:
+
+```
+VITE_SASQUATCH_AI_BASE=https://ai.skate-spots.org/api/ai
+```
+
+Leave it unset and the app keeps using `/api/ai` (the Netlify Function), so the
+move is reversible and can be rolled out gradually. The proxy's CORS allowlist
+already includes `https://skate-spots.org` and the local/Capacitor origins.
+
+## Verify
+
+```bash
+# from anywhere, once TLS is up:
+curl -s -X POST https://ai.skate-spots.org/api/ai/status
+# → {"configured":true,...}   (owner balance; public on purpose)
+```
+
+## Notes
+
+- `SASQUATCH_REQUIRE_PRO=1` keeps the Pro gate (owner pubkeys bypass). A client
+  without a valid NIP-98 header gets 401; without an active Pro pass, 402.
+- The proxy is per-IP throttled at 20 req/min. It's a single long-lived process,
+  so the throttle actually holds (unlike the serverless version).
+- `netlify/functions/ai.mjs` stays in the repo as the fallback path; nothing
+  breaks if this service is down and the client is still pointed at `/api/ai`.
--- a/ai-proxy/package.json
+++ b/ai-proxy/package.json
@ -0,0 +1,12 @@
+{
+  "name": "skatespots-ai-proxy",
+  "private": true,
+  "type": "module",
+  "scripts": {
+    "start": "bun run server.ts",
+    "dev": "bun --watch run server.ts"
+  },
+  "dependencies": {
+    "nostr-tools": "^2.13.0"
+  }
+}
--- a/ai-proxy/server.ts
+++ b/ai-proxy/server.ts
@ -0,0 +1,282 @@
+// SkateSpots AI proxy — runs on the CVM host (Bun), behind Caddy for TLS.
+//
+// This is the same job the Netlify Function `netlify/functions/ai.mjs` does, moved
+// onto infrastructure we own so a user-facing AI call no longer spends Netlify
+// compute/bandwidth credits — and so the sats-paid Routstr key lives on our box,
+// not a third party's.
+//
+//   • POST /api/ai        — chat proxy (curated model allowlist; optional Pro gate)
+//   • POST /api/ai/status — owner-only upstream balance (low-balance nudge)
+//
+// The upstream is provider-agnostic. For sats-paid inference it points at a
+// Routstr node (SASQUATCH_AI_BASE + SASQUATCH_AI_KEY).
+import { verifyEvent } from 'nostr-tools';
+
+// Bun global (typed minimally so this file needs no bun-types dependency).
+declare const Bun: {
+  serve(options: {
+    port: number;
+    hostname?: string;
+    fetch: (request: Request) => Response | Promise<Response>;
+  }): { stop: () => void };
+};
+
+const PORT = Number(process.env.AI_PROXY_PORT || 8787);
+const HOSTNAME = process.env.AI_PROXY_HOST || '127.0.0.1';
+
+const AI_BASE = (
+  process.env.SASQUATCH_AI_BASE ||
+  process.env.OPENROUTER_BASE_URL ||
+  'https://openrouter.ai/api/v1'
+).replace(/\/+$/, '');
+const AI_URL = `${AI_BASE}/chat/completions`;
+const AI_KEY = process.env.SASQUATCH_AI_KEY || process.env.OPENROUTER_API_KEY;
+
+const DEFAULT_MODELS = [
+  'meta-llama/llama-3.3-70b-instruct',
+  'deepseek/deepseek-chat-v3-0324',
+  'x-ai/grok-4.5',
+  'moonshotai/kimi-k2',
+  'qwen/qwen-2.5-72b-instruct',
+  'mistralai/mistral-nemo',
+  'meta-llama/llama-3.1-8b-instruct',
+];
+const ALLOWED_MODELS = new Set(
+  (process.env.SASQUATCH_MODELS || DEFAULT_MODELS.join(','))
+    .split(',')
+    .map((s) => s.trim())
+    .filter(Boolean),
+);
+
+// ── Pro enforcement ───────────────────────────────────────────────────────────
+const REQUIRE_PRO = process.env.SASQUATCH_REQUIRE_PRO === '1';
+const PLATFORM_PRO_ADDRESS = 'skatespots@minibits.cash';
+/** Primary platform pubkey — zap requests must tag it; owners always pass. */
+const PLATFORM_PUBKEY = '20e7c953439e66f3882ab96fe0faa8ba246d6fb12c03b0cb4e2dc4623b8bcb21';
+const PLATFORM_OWNER_PUBKEYS = new Set([
+  PLATFORM_PUBKEY,
+  'caba497f7beda0650d6b78ab211216e4d2ebabeec9ca6eb2a05cee8e55c6871b',
+]);
+const DAY_SECONDS = 86_400;
+/** Tier days ↔ sats. Keep in sync with PRO_TIERS in src/lib/sasquatchPro.ts. */
+const PRO_TIERS = [
+  { days: 7, sats: 2000 },
+  { days: 30, sats: 5000 },
+  { days: 90, sats: 12000 },
+  { days: 365, sats: 36000 },
+];
+
+let providerPubkeyCache: string | null | undefined;
+async function getProProviderPubkey(): Promise<string | null> {
+  if (providerPubkeyCache !== undefined) return providerPubkeyCache;
+  try {
+    const [user, domain] = PLATFORM_PRO_ADDRESS.split('@');
+    const res = await fetch(`https://${domain}/.well-known/lnurlp/${user}`, {
+      signal: AbortSignal.timeout(6000),
+    });
+    const j = (await res.json()) as { nostrPubkey?: string };
+    providerPubkeyCache = j.nostrPubkey || null;
+  } catch {
+    providerPubkeyCache = null;
+  }
+  return providerPubkeyCache;
+}
+
+function decodeB64Json(value: string) {
+  try {
+    return JSON.parse(Buffer.from(value, 'base64').toString('utf8'));
+  } catch {
+    return null;
+  }
+}
+function tagVal(tags: string[][] | undefined, name: string): string | undefined {
+  return tags?.find((t) => t[0] === name)?.[1];
+}
+function satsFromBolt11(pr: string): number {
+  const m = /^lnbc(\d+)([munp]?)/i.exec(String(pr).toLowerCase());
+  if (!m) return 0;
+  const n = Number(m[1]);
+  const mult = { m: 1e-3, u: 1e-6, n: 1e-9, p: 1e-12 }[m[2] as 'm' | 'u' | 'n' | 'p'] ?? 1;
+  return Math.round(n * mult * 1e8);
+}
+
+/** Verify a NIP-98 auth header → pubkey, or null. */
+function verifyNip98(authHeader: string | null): string | null {
+  if (!authHeader) return null;
+  const ev = decodeB64Json(authHeader);
+  if (!ev || ev.kind !== 27235) return null;
+  const method = tagVal(ev.tags, 'method');
+  if (method && method !== 'POST') return null;
+  const u = tagVal(ev.tags, 'u') ?? '';
+  if (u && !u.endsWith('/api/ai')) return null;
+  if (Math.abs(Date.now() / 1000 - ev.created_at) > 12 * 3600) return null;
+  try {
+    return verifyEvent(ev) ? ev.pubkey : null;
+  } catch {
+    return null;
+  }
+}
+
+/** Verify a Pro pass (provider-authored zap receipt bound to the caller). */
+async function verifyPro(receiptHeader: string | null, userPubkey: string): Promise<number | null> {
+  if (!receiptHeader) return null;
+  const receipt = decodeB64Json(receiptHeader);
+  if (!receipt || receipt.kind !== 9735) return null;
+  const providerPubkey = await getProProviderPubkey();
+  if (!providerPubkey || receipt.pubkey !== providerPubkey) return null;
+  try {
+    if (!verifyEvent(receipt)) return null;
+  } catch {
+    return null;
+  }
+  const desc = tagVal(receipt.tags, 'description');
+  if (!desc) return null;
+  let zapReq;
+  try {
+    zapReq = JSON.parse(desc);
+  } catch {
+    return null;
+  }
+  if (!zapReq || zapReq.pubkey !== userPubkey) return null;
+  try {
+    if (!verifyEvent(zapReq)) return null;
+  } catch {
+    return null;
+  }
+  if (!zapReq.tags?.some((t: string[]) => t[0] === 'p' && t[1] === PLATFORM_PUBKEY)) return null;
+
+  let paidSats = satsFromBolt11(tagVal(receipt.tags, 'bolt11') ?? '');
+  if (!paidSats) {
+    const msat = Number(tagVal(zapReq.tags, 'amount'));
+    if (Number.isFinite(msat) && msat > 0) paidSats = Math.floor(msat / 1000);
+  }
+  const tier = [...PRO_TIERS].sort((a, b) => b.sats - a.sats).find((t) => paidSats >= t.sats);
+  if (!tier) return null;
+  return receipt.created_at + tier.days * DAY_SECONDS;
+}
+
+// Origins allowed to call the proxy (web + local dev + Capacitor native).
+const ALLOWED_ORIGINS = new Set([
+  'https://skate-spots.org',
+  'http://localhost:8080',
+  'http://localhost:5173',
+  'capacitor://localhost',
+  'http://localhost',
+  'https://localhost',
+]);
+function corsHeaders(origin: string): Record<string, string> {
+  return {
+    'Access-Control-Allow-Origin': ALLOWED_ORIGINS.has(origin) ? origin : 'https://skate-spots.org',
+    'Access-Control-Allow-Methods': 'POST, OPTIONS',
+    'Access-Control-Allow-Headers': 'Content-Type, x-sasquatch-auth, x-sasquatch-pro',
+    Vary: 'Origin',
+  };
+}
+
+function json(body: unknown, status: number, headers: Record<string, string>): Response {
+  return new Response(JSON.stringify(body), {
+    status,
+    headers: { ...headers, 'Content-Type': 'application/json' },
+  });
+}
+
+// Per-IP throttle. A single long-lived process (unlike serverless), so this is a
+// more meaningful damper than the Netlify version.
+const WINDOW_MS = 60_000;
+const LIMIT = 20;
+const hits = new Map<string, number[]>();
+
+async function handleStatus(headers: Record<string, string>): Promise<Response> {
+  if (!AI_KEY) return json({ configured: false }, 200, headers);
+  try {
+    const r = await fetch(`${AI_BASE}/balance/info`, {
+      headers: { Authorization: `Bearer ${AI_KEY}` },
+      signal: AbortSignal.timeout(8000),
+    });
+    const j = (await r.json().catch(() => ({}))) as { balance?: number };
+    const balanceSats = Math.floor(Number(j.balance ?? 0) / 1000);
+    return json({ configured: true, balanceSats, low: balanceSats < 500, provider: AI_BASE }, 200, headers);
+  } catch {
+    return json({ configured: true, balanceSats: null, low: false }, 200, headers);
+  }
+}
+
+async function handleChat(request: Request, headers: Record<string, string>): Promise<Response> {
+  if (REQUIRE_PRO) {
+    const pubkey = verifyNip98(request.headers.get('x-sasquatch-auth'));
+    if (!pubkey) return json({ error: 'auth required' }, 401, headers);
+    if (!PLATFORM_OWNER_PUBKEYS.has(pubkey)) {
+      const proUntil = await verifyPro(request.headers.get('x-sasquatch-pro'), pubkey);
+      if (!proUntil || proUntil < Math.floor(Date.now() / 1000)) {
+        return json({ error: 'pro required' }, 402, headers);
+      }
+    }
+  }
+
+  const key = AI_KEY;
+  if (!key) return json({ error: 'AI proxy not configured' }, 503, headers);
+
+  const ip =
+    request.headers.get('x-forwarded-for')?.split(',')[0].trim() ||
+    request.headers.get('x-real-ip') ||
+    'unknown';
+  const now = Date.now();
+  const recent = (hits.get(ip) ?? []).filter((t) => now - t < WINDOW_MS);
+  if (recent.length >= LIMIT) return json({ error: 'Too many requests' }, 429, headers);
+  recent.push(now);
+  hits.set(ip, recent);
+
+  let body: { model?: string; messages?: unknown[]; max_tokens?: number; stream?: boolean };
+  try {
+    body = (await request.json()) as typeof body;
+  } catch {
+    return json({ error: 'Invalid JSON' }, 400, headers);
+  }
+
+  const requested = String(body.model ?? '');
+  const model = ALLOWED_MODELS.has(requested) ? requested : ALLOWED_MODELS.values().next().value;
+  const messages = Array.isArray(body.messages) ? body.messages.slice(-20) : [];
+  const maxTokens = Math.min(Number(body.max_tokens) || 400, 800);
+  const stream = body.stream === true;
+
+  let upstream: Response;
+  try {
+    upstream = await fetch(AI_URL, {
+      method: 'POST',
+      headers: {
+        Authorization: `Bearer ${key}`,
+        'Content-Type': 'application/json',
+        'HTTP-Referer': 'https://skate-spots.org',
+        'X-Title': 'SkateSpots',
+      },
+      body: JSON.stringify({ model, messages, max_tokens: maxTokens, stream }),
+      signal: AbortSignal.timeout(30_000),
+    });
+  } catch (err) {
+    return json({ error: `Upstream failed: ${String(err)}` }, 502, headers);
+  }
+
+  const passthrough = { ...headers };
+  const contentType = upstream.headers.get('content-type');
+  if (contentType) passthrough['Content-Type'] = contentType;
+  return new Response(upstream.body, { status: upstream.status, headers: passthrough });
+}
+
+Bun.serve({
+  port: PORT,
+  hostname: HOSTNAME,
+  async fetch(request: Request): Promise<Response> {
+    const origin = request.headers.get('origin') || '';
+    const headers = corsHeaders(origin);
+    const path = new URL(request.url).pathname;
+
+    if (request.method === 'OPTIONS') return new Response(null, { status: 204, headers });
+    if (request.method !== 'POST') return json({ error: 'Method not allowed' }, 405, headers);
+
+    if (path === '/api/ai/status') return handleStatus(headers);
+    if (path === '/api/ai') return handleChat(request, headers);
+    return json({ error: 'Not found' }, 404, headers);
+  },
+});
+
+console.log(`[ai-proxy] listening on http://${HOSTNAME}:${PORT} (upstream ${AI_BASE})`);
--- a/ai-proxy/skatespots-ai.service
+++ b/ai-proxy/skatespots-ai.service
@ -0,0 +1,28 @@
+# ─────────────────────────────────────────────────────────────────────────────
+# systemd unit for the SkateSpots AI proxy (Bun, on the CVM host).
+# Installed by scripts/deploy-ai-proxy.sh — do not hand-edit on the host.
+# ─────────────────────────────────────────────────────────────────────────────
+
+[Unit]
+Description=SkateSpots AI proxy (Routstr)
+After=network-online.target
+Wants=network-online.target
+
+[Service]
+Type=simple
+User=ubuntu
+WorkingDirectory=/home/ubuntu/skatespots-ai
+EnvironmentFile=/home/ubuntu/skatespots-ai/.env
+ExecStart=/home/ubuntu/.bun/bin/bun run server.ts
+Restart=always
+RestartSec=5
+StandardOutput=journal
+StandardError=journal
+SyslogIdentifier=skatespots-ai
+
+# Basic hardening — mirrors the CVM unit (bun stays happy; binds localhost only).
+NoNewPrivileges=true
+PrivateTmp=true
+
+[Install]
+WantedBy=multi-user.target
--- a/ai-proxy/tsconfig.json
+++ b/ai-proxy/tsconfig.json
@ -0,0 +1,15 @@
+{
+  // Type-checks the standalone Bun AI proxy. Separate from the app's root
+  // tsconfig (which only includes `src`) and run by `npm test` via `typecheck:ai`.
+  "compilerOptions": {
+    "target": "ES2022",
+    "module": "ESNext",
+    "moduleResolution": "bundler",
+    "lib": ["ESNext"],
+    "types": ["node"],
+    "esModuleInterop": true,
+    "skipLibCheck": true,
+    "noEmit": true
+  },
+  "include": ["server.ts"]
+}
--- a/eslint.config.js
+++ b/eslint.config.js
@ -22,6 +22,8 @@ export default tseslint.config(
      // `typecheck:cvm` script, wired into `npm test`), not by the web ESLint
      // config (it leans on `any` and has no React rules).
      "cvm-server",
+      // Standalone Bun AI proxy — same deal: own tsconfig (`typecheck:ai`).
+      "ai-proxy",
      // Static utility HTML (icon generators, sticker sheets, QR pages) that
      // are intentionally minimal and don't need app-level meta/OG tags.
      "public/**/*.html",
--- a/package.json
+++ b/package.json
@ -8,9 +8,12 @@
    "dev": "npm i --silent && vite",
    "build": "npm i --silent && vite build -l error && cp dist/index.html dist/404.html && echo 'Project built successfully!'",
    "deploy:cvm": "bash scripts/deploy-cvm.sh",
+    "deploy:ai": "bash scripts/deploy-ai-proxy.sh",
    "typecheck:cvm": "tsc -p cvm-server/tsconfig.json",
+    "typecheck:ai": "tsc -p ai-proxy/tsconfig.json",
    "check:cvm": "bun build cvm-server/server.ts --target=bun --outfile=/tmp/cvm-check.js",
-    "test": "npm i --silent && tsc --noEmit && tsc -p cvm-server/tsconfig.json && eslint && vitest run --reporter=dot --silent && vite build -l error && cp dist/index.html dist/404.html && echo 'All tests passed!'",
+    "check:ai": "bun build ai-proxy/server.ts --target=bun --outfile=/tmp/ai-check.js",
+    "test": "npm i --silent && tsc --noEmit && tsc -p cvm-server/tsconfig.json && tsc -p ai-proxy/tsconfig.json && eslint && vitest run --reporter=dot --silent && vite build -l error && cp dist/index.html dist/404.html && echo 'All tests passed!'",
    "deploy": "npm i --silent && npm run build && npx -y nostr-deploy-cli deploy --skip-setup",
    "android:init": "npm run android:build:apk && npx cap add android && npx cap sync android && node scripts/inject-amber-queries.mjs && node scripts/inject-main-activity.mjs && node scripts/inject-network-security.mjs && node scripts/inject-signing-config.mjs",
    "android:sync": "npm run android:build:apk && npx cap sync android && node scripts/sync-android-version.mjs && node scripts/inject-amber-queries.mjs && node scripts/inject-main-activity.mjs && node scripts/inject-network-security.mjs && node scripts/inject-signing-config.mjs",
--- a/scripts/deploy-ai-proxy.sh
+++ b/scripts/deploy-ai-proxy.sh
@ -0,0 +1,59 @@
+#!/usr/bin/env bash
+#
+# Deploy the SkateSpots AI proxy to the CVM host (same box as the CVM server).
+#
+# It is the same class of service as deploy-cvm.sh: a Bun process under systemd,
+# reachable only via Caddy (TLS) on the public side. Config/host come from
+# ~/.config/skatespots/cvm.env (never committed).
+#
+#   npm run deploy:ai        # or: bash scripts/deploy-ai-proxy.sh
+#   bash scripts/deploy-ai-proxy.sh --logs
+set -euo pipefail
+
+REPO_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+
+CONF="${HOME}/.config/skatespots/cvm.env"
+if [[ -f "$CONF" ]]; then
+  # shellcheck disable=SC1090
+  source "$CONF"
+fi
+
+: "${CVM_HOST:?Set CVM_HOST (e.g. in ~/.config/skatespots/cvm.env) — the host is never committed}"
+CVM_SSH_KEY="${CVM_SSH_KEY:-$HOME/.ssh/skatespots-new}"
+AI_DIR="${AI_DIR:-/home/ubuntu/skatespots-ai}"
+AI_SERVICE="${AI_SERVICE:-skatespots-ai}"
+
+SSH=(ssh -i "$CVM_SSH_KEY" -o BatchMode=yes -o ConnectTimeout=12 "$CVM_HOST")
+
+if [[ "${1:-}" == "--logs" ]]; then
+  "${SSH[@]}" "journalctl -u $AI_SERVICE -n 40 --no-pager"
+  exit 0
+fi
+
+echo "→ Ensuring $AI_DIR exists"
+"${SSH[@]}" "mkdir -p $AI_DIR"
+
+echo "→ Uploading ai-proxy/{server.ts,package.json,skatespots-ai.service}"
+scp -i "$CVM_SSH_KEY" -o BatchMode=yes \
+  "$REPO_ROOT/ai-proxy/server.ts" \
+  "$REPO_ROOT/ai-proxy/package.json" \
+  "$REPO_ROOT/ai-proxy/skatespots-ai.service" \
+  "$CVM_HOST:$AI_DIR/"
+
+echo "→ bun install"
+"${SSH[@]}" "cd $AI_DIR && ~/.bun/bin/bun install --silent 2>&1 | tail -3 || true"
+
+echo "→ Installing the unit and (re)starting $AI_SERVICE"
+"${SSH[@]}" "sudo cp $AI_DIR/skatespots-ai.service /etc/systemd/system/ \
+  && sudo systemctl daemon-reload \
+  && sudo systemctl enable --now $AI_SERVICE >/dev/null 2>&1 || true; \
+  sudo systemctl restart $AI_SERVICE \
+  && sleep 2 && echo \"active: \$(systemctl is-active $AI_SERVICE)\""
+
+echo "→ Local health check (through the service, on the host)"
+"${SSH[@]}" "curl -s -o /dev/null -w 'status endpoint: %{http_code}\n' -X POST http://127.0.0.1:8787/api/ai/status || true"
+
+echo "→ Recent log:"
+"${SSH[@]}" "journalctl -u $AI_SERVICE -n 15 --no-pager | tail -6"
+
+echo "✅ AI proxy deployed. Public TLS is served by Caddy — see ai-proxy/DEPLOY.md."
--- a/src/hooks/useSasquatchAiStatus.ts
+++ b/src/hooks/useSasquatchAiStatus.ts
@ -1,7 +1,7 @@
 import { useQuery } from '@tanstack/react-query';
-import { Capacitor } from '@capacitor/core';
 import { useCurrentUser } from '@/hooks/useCurrentUser';
 import { PLATFORM_OWNER_PUBKEYS } from '@/hooks/useListingRestrictions';
+import { sasquatchAiStatusUrl } from '@/lib/sasquatchAi';

 /** Report shape from the proxy's owner-only `/api/ai/status`. */
 export interface SasquatchAiStatus {
@ -11,9 +11,7 @@ export interface SasquatchAiStatus {
  provider?: string;
 }

-const AI_STATUS_URL = Capacitor.isNativePlatform()
-  ? 'https://skate-spots.org/api/ai/status'
-  : '/api/ai/status';
+const AI_STATUS_URL = sasquatchAiStatusUrl();

 /**
 * Owner-only: the Sasquatch AI (upstream) account balance, so the operator can be
--- a/src/lib/proAuth.ts
+++ b/src/lib/proAuth.ts
@ -10,6 +10,7 @@
 * signer (Amber) doesn't pop on every message.
 */
 import type { NostrEvent } from '@nostrify/nostrify';
+import { sasquatchAiBase } from '@/lib/sasquatchAi';

 type SignEvent = (event: {
  kind: number;
@ -40,7 +41,7 @@ export async function buildProAuthHeaders(opts: {
      kind: 27235,
      created_at: Math.floor(Date.now() / 1000),
      tags: [
-        ['u', 'https://skate-spots.org/api/ai'],
+        ['u', sasquatchAiBase()],
        ['method', 'POST'],
      ],
      content: '',
--- a/src/lib/sasquatchAi.ts
+++ b/src/lib/sasquatchAi.ts
@ -0,0 +1,30 @@
+import { isNativePlatform } from '@/lib/platform';
+
+/**
+ * Base URL of the Sasquatch AI proxy — the single place the app decides where AI
+ * requests go.
+ *
+ * Default: the original Netlify Function at `/api/ai` (same-origin on web; the
+ * absolute production origin on native, which has no same-origin server).
+ *
+ * To move AI off Netlify onto our own host (ai-proxy/ on the CVM host), set the
+ * build-time override:
+ *   VITE_SASQUATCH_AI_BASE=https://ai.skate-spots.org/api/ai
+ * This is a pure routing switch and trivially reversible: the proxy accepts any
+ * host — it only checks that a NIP-98 `u` value ends in `/api/ai`.
+ *
+ * Returns an absolute URL, so it's safe to use both for `fetch()` and as the
+ * NIP-98 `u` tag.
+ */
+export function sasquatchAiBase(): string {
+  const override = (import.meta.env as Record<string, string | undefined>).VITE_SASQUATCH_AI_BASE;
+  if (override) return override.replace(/\/+$/, '');
+  if (isNativePlatform()) return 'https://skate-spots.org/api/ai';
+  const origin = typeof location !== 'undefined' ? location.origin : 'https://skate-spots.org';
+  return `${origin}/api/ai`;
+}
+
+/** Owner balance endpoint (`…/api/ai/status`). */
+export function sasquatchAiStatusUrl(): string {
+  return `${sasquatchAiBase()}/status`;
+}
--- a/src/lib/sasquatchBrain.ts
+++ b/src/lib/sasquatchBrain.ts
@ -4,20 +4,16 @@
 * surface can reuse the same model list + fallback behaviour.
 */

-import { Capacitor } from '@capacitor/core';
+import { sasquatchAiBase } from '@/lib/sasquatchAi';

 // ─── AI models, via our own server-side proxy ────────────────────────────────
-// SECURITY: no API key is ever shipped to the client. Requests go to our Netlify
-// Function at `/api/ai`, which runs against Netlify's AI Gateway server-side and
-// only allows a curated model list (see netlify/functions/ai.mjs). If the proxy
-// is unavailable we fall back to the keyless Pollinations provider below, so
-// Sasquatch still works with zero secrets shipped.
-//
-// The native (Capacitor) build has no same-origin `/api/ai`, so it calls the
-// absolute production URL; the web build stays same-origin.
-const AI_PROXY_URL = Capacitor.isNativePlatform()
-  ? 'https://skate-spots.org/api/ai'
-  : '/api/ai';
+// SECURITY: no API key is ever shipped to the client. Requests go to our proxy
+// (`/api/ai`), which holds the Routstr key server-side and only allows a curated
+// model list. Where that proxy lives is decided in one place — lib/sasquatchAi.ts
+// (the Netlify Function by default, or our own ai-proxy/ service on the CVM host
+// via VITE_SASQUATCH_AI_BASE). If it's unavailable we fall back to the keyless
+// Pollinations provider below, so Sasquatch still works with zero secrets shipped.
+const AI_PROXY_URL = sasquatchAiBase();

 // Models Sasquatch requests through the proxy, in preference order. These are
 // the slugs served by our upstream (a Routstr node — see AGENTS.md), which are