Adrien

This is the Adrien dashboard for Juliano Pizzaria, a Next.js app.

Getting Started

npm install
npm run dev

Open http://localhost:3000 with your browser to see the result.

Password protection

The whole app is gated behind a single shared password (no per-user accounts yet) via middleware.ts. Set DASHBOARD_PASSWORD in your environment (and in Vercel for production) — without it, the app is left open so a fresh unconfigured environment never accidentally locks out its own developer. The password is checked against a hash stored in an httpOnly cookie, not sent back and forth on every request.

Data providers

Each data source lives behind a small provider interface in lib/providers/, so live integrations can be swapped in without touching the dashboard UI. Provider status:

Source	Status	Env vars
Meta (Facebook/Instagram) Ads	Live (falls back to mock if unconfigured)	META_ACCESS_TOKEN, META_AD_ACCOUNT_ID
Google Ads	Code ready, needs credentials	GOOGLE_ADS_CLIENT_ID, GOOGLE_ADS_CLIENT_SECRET, GOOGLE_ADS_REFRESH_TOKEN, GOOGLE_ADS_CUSTOMER_ID, GOOGLE_ADS_LOGIN_CUSTOMER_ID (optional), GOOGLE_ADS_DEVELOPER_TOKEN (optional, ignored by Google as of Sept 2026)
Bite Business	Blocked — see below	—
Social (Facebook/Instagram/TikTok/YouTube)	Mock	—

Meta Ads setup

1. Create a Meta app with Marketing API access and generate a long-lived access token with ads_read permission for the Juliano Pizzaria ad account.
2. Set META_ACCESS_TOKEN in your environment (and in Vercel Project Settings → Environment Variables for production).
3. Optionally set META_AD_ACCOUNT_ID — defaults to Juliano Pizzaria's ad account (937731679434253).
4. Without a token, /api/ads transparently serves mock data for Meta so the app never breaks in an unconfigured environment — each ad platform card shows a "Live data" or "Mock data" badge so it's clear which is which.

Note: Juliano Pizzaria's Meta ad account currently has no purchase/conversion tracking configured (campaigns are optimizing for reach, not purchases), so revenue, ROAS, conversions, and CPA show as "Not tracked" until a pixel or Conversions API event is set up and campaigns switch to a purchase objective. Spend, impressions, clicks, reach, CTR, and CPC are live today.

Google Ads setup

Google changed this on September 10, 2026: developer tokens are sunset — access level is now tied to the Google Cloud project behind your OAuth client, not a separate token. GOOGLE_ADS_DEVELOPER_TOKEN is optional here (kept only in case Google reinstates the header check in a future API version).

1. Create a Google Cloud project and enable the Google Ads API on it (APIs & Services → Library → search "Google Ads API" → Enable). This gives the project Test access by default, which only works against test accounts.
2. From that same project's Google Ads API Overview page in Cloud Console, apply for Basic access (needed to query Juliano Pizzaria's real account) — Google now gates this behind brand verification.
3. Create an OAuth 2.0 Client ID (type: Web application) on that project, with https://developers.google.com/oauthplayground as an authorized redirect URI.
4. Generate a refresh token using Google's OAuth Playground: gear icon → "Use your own OAuth credentials" → paste Client ID/Secret → authorize the https://www.googleapis.com/auth/adwords scope → exchange the code.
5. Find the customer ID (the number shown top-right in Google Ads, no dashes) for Juliano Pizzaria's account. If it's managed under an agency/MCC account (the developer token application itself is applied from the MCC), also set GOOGLE_ADS_LOGIN_CUSTOMER_ID to the MCC's customer ID.
6. Set GOOGLE_ADS_CLIENT_ID, GOOGLE_ADS_CLIENT_SECRET, GOOGLE_ADS_REFRESH_TOKEN, GOOGLE_ADS_CUSTOMER_ID (and optionally GOOGLE_ADS_LOGIN_CUSTOMER_ID) in Vercel.
7. Without them, /api/ads transparently serves mock data for Google, same as Meta — the card's badge and an inline message name exactly which env var is missing.

Bite Business setup — currently blocked

documentation.getbite.com documents a different product (a kiosk/online-ordering platform) from the one Juliano Pizzaria actually uses (bitebusiness.com, a delivery-management platform — own drivers or Uber Direct). That doc's API isn't applicable here.

More importantly: Bite Business's own pricing page shows API access is only included on "The Lot" plan ($99/week); Juliano's account is on "Main" ($59/week), which doesn't include API access at all. So this integration is blocked regardless of which docs are correct, until either:

1. The account is upgraded to "The Lot," and Bite Business support is asked for API docs/credentials for that plan tier, or
2. Bite Business offers another way to get sales/order data out (a CSV export, webhook, or similar) that doesn't require the API-access plan feature.

Business Data stays on mock data until one of those is resolved.

Creative Factory

A separate tab (/creative-factory) with two pieces:

1. Ad copy generator (/api/creative/copy) — writes headline/primary text/ description/CTA variations for a product/offer, plus a free-form "extra details" brief (must-include phrases, brand guidelines, etc.) that's folded into the prompt. Uses the Anthropic API for genuinely varied copy when ANTHROPIC_API_KEY is set; otherwise falls back to fixed templates so it always returns something usable.
2. Meta campaign builder & launcher (/api/creative/launch, /api/creative/activate) — builds a full campaign → ad set → ad creative → ad directly via the Graph API, with:
   • Full audience targeting: countries, age range, gender, and free-text interests (resolved to Meta's interest-targeting IDs via the ad interest search endpoint — best match per name, unmatched names are skipped rather than failing the request).
   • A Conversions objective with pixel ID + conversion event fields (Juliano's Meta account has no pixel configured yet, so this objective will fail until one is set up — the UI flags this).
   • Up to two images — a 1:1 square and a 9:16 vertical — with live aspect-ratio previews in the form. When both are provided, the creative uses asset_feed_spec so Meta auto-selects the right image per placement (feed vs. Stories/Reels) instead of stretching one image everywhere.
   • A review/preview step (mock ad card + a plain-English summary of the campaign/ad set/audience) that has to be explicitly confirmed before anything is sent to Meta at all.

Safety design — always creates paused, never auto-activates: every campaign, ad set, and ad is created with status: PAUSED. Nothing spends money on creation. The UI shows the draft with a link to review it in Meta Ads Manager, and a separate, explicit "Activate" button that's the only thing that flips it live — that action is never taken automatically.

Setup: needs a META_ACCESS_TOKEN with ads_management permission (the reporting integration only needed ads_read — this is a different, broader scope, so the existing token will need to be regenerated), plus META_PAGE_ID — the Facebook Page ID that ads run from (find it under the Page's About section, or via Business Settings → Accounts → Pages).

3. Google campaign builder & launcher (/api/creative/google/copy, /api/creative/google/launch, /api/creative/google/activate) — a second platform tab alongside Meta, building Search, Display, or Performance Max campaigns directly via the Google Ads API:
   • Campaign → one or more ad groups (Search/Display) or asset groups (Performance Max), each with its own headlines (≤30 chars), descriptions (≤90 chars), and — for Display/Performance Max — one or more images.
   • Search ad groups take keywords typed one per line, using the same syntax as Google Ads Editor: plain text = broad match, "in quotes" = phrase match, [in brackets] = exact match.
   • A "Generate headlines, descriptions & keywords" button per ad group calls the Anthropic API (same product/offer/tone/details brief as the Meta copy generator) for AI-written Google-length copy plus keyword ideas, falling back to templates without ANTHROPIC_API_KEY.
   • Performance Max's asset requirements are approximated (Google recommends more headlines/images than the bare minimum used here) — first drafts may need small tweaks in the Google Ads UI before they clear review, flagged in the UI itself.

Same safety design as Meta: campaigns, ad groups/asset groups, and ads are always created PAUSED. A separate "Activate" button is the only thing that enables them — never automatic.

Setup: reuses the same GOOGLE_ADS_CLIENT_ID, GOOGLE_ADS_CLIENT_SECRET, GOOGLE_ADS_REFRESH_TOKEN, GOOGLE_ADS_CUSTOMER_ID credentials as the read-only reporting integration above — no extra setup needed as long as the OAuth token was authorized with the https://www.googleapis.com/auth/adwords scope (which grants both read and write).

Content Studio & Scheduler

Two more tabs, for repurposing existing photos/video (or generating from scratch) across every social platform and queueing it to go out:

1. Content Studio (/api/content/repurpose) — upload a photo or video, generate one with AI (/api/content/generate-image — OpenAI's image API, from the same product/offer/tone/details brief as the copy generators), or leave it blank entirely, pick platforms (Meta, YouTube, LinkedIn, TikTok, Twitter/X), and it generates a tailored caption + hashtags per platform (Anthropic API, template fallback) plus a center-cropped image resized to each platform's native aspect ratio (via sharp).
   • Video is now resized (crop/scale to each platform's aspect ratio, trimmed to its max length) via ffmpeg (fluent-ffmpeg + @ffmpeg-installer/ffmpeg), with the result uploaded to Vercel Blob and scheduled by URL rather than inline bytes. This is genuinely best-effort — ffmpeg in a serverless function has real constraints (execution time limits, /tmp size, no persistent disk between invocations) that couldn't be fully validated outside an actual Vercel deploy; the pipeline itself was tested end-to-end locally (a real ffmpeg crop/scale/trim run) but not against Vercel's specific limits. If it fails, the platform falls back to a note with the target spec instead of a broken file. Requires BLOB_READ_WRITE_TOKEN (see Setup below) — without it, video falls back to the target-spec note as before.
   • AI video generation (Runway/Luma/Veo/Kling) is still not wired up — each is a separate paid API, a further fast-follow once one's chosen.
2. Scheduler (/api/content/schedule, /api/content/publish-due) — a queue of everything scheduled from Content Studio. A Vercel Cron job (vercel.json, daily — see the Cron note below) hits /api/content/publish-due:
   • Meta posts automatically via the Facebook Page's /photos or /feed endpoint (an organic post, separate from the paid ad campaigns above) when a post comes due.
   • YouTube, LinkedIn, TikTok, Twitter/X don't have posting APIs wired up yet — each needs its own developer app, OAuth consent, and (for TikTok in particular) a manual approval process. Due posts on these platforms flip to "needs manual post" in the Scheduler so nothing silently fails to go out — you post it yourself and click "Mark posted."

Setup:

• Meta auto-posting needs META_ACCESS_TOKEN with pages_manage_posts permission (a third scope, distinct from ads_read and ads_management) plus the existing META_PAGE_ID.
• The scheduler queue needs UPSTASH_REDIS_REST_URL and UPSTASH_REDIS_REST_TOKEN (add a Redis integration from the Vercel Marketplace) for persistence. Without it, the queue falls back to in-memory storage that doesn't survive a restart or cold start — fine for trying it out, not for production use.
• Video (resizing, or storing one in Adrien Brain's content library) needs BLOB_READ_WRITE_TOKEN (add a Blob store from the Vercel project's Storage tab, which sets this automatically). Redis's REST API caps value sizes well below what a video file needs, so Blob is used for video specifically — images still go through Redis/in-memory as before.
• Vercel Cron note: the Hobby plan only allows daily cron runs, so vercel.json schedules /api/content/publish-due once a day (9am UTC). A more frequent check (every 5–15 min, so scheduled posts go out closer to their target time) needs a Pro plan — tighten the schedule there once you're on one. A 5-minute schedule on Hobby isn't just throttled, it blocks every deployment outright (Vercel rejects the whole build at validation), so don't loosen this without upgrading first.

Adrien Brain

The orchestration layer on top of Content Studio and the Scheduler — the goal is simple: never let a platform go more than a day without a post, two at the absolute most.

• Content library (/api/content/library) — a standing folder of photos to repurpose, separate from any one Content Studio session. Add to it any time; each asset tracks how many times it's been used so Adrien Brain favors whatever's been sitting the longest.
• Brand brief (/api/content/brand) — a saved default product/offer/tone/details brief, so Adrien Brain doesn't need you to retype one every time it proposes something on its own.
• The plan (/api/content/plan) — on load, checks every platform's most recent scheduled/published/manual-post entry in the Scheduler. Any platform within 4 hours of the 24h mark shows as "Due soon"; past 48h shows as "Overdue." For each, it proposes a time, a caption (generated from the brand brief), and either the least-recently-used library asset ("repurpose") or a "Generate an image" button when the library's empty or exhausted. One click — "Approve & schedule" — sends it straight into the same Scheduler queue everything else uses.
• Video can now be added to the library too (needs BLOB_READ_WRITE_TOKEN — see above), stored and previewed by URL rather than inline like images.
• 2-week idea calendar (/api/content/calendar) — one flagship content idea per day for the next 14 days (format: reel or static; tone: funny, serious, warm, or informative), adapt it per platform when you actually schedule it. Informed by real engagement/follower-trend numbers pulled from the Social Media Review data where available (Meta ← Facebook + Instagram, YouTube, TikTok — LinkedIn and Twitter/X aren't tracked there yet, so those get no performance signal). There's no live trends feed integrated (no TikTok Trends API, no Google Trends) — the AI reasons from its own general knowledge of platform trends instead, and says so in its reasoning rather than presenting it as live data. Template fallback without ANTHROPIC_API_KEY.

This reuses the Scheduler's Redis/in-memory backend and Content Studio's image generation — no new env vars beyond what those already need, plus BLOB_READ_WRITE_TOKEN for video.

Deploy on Vercel

Import this repository in Vercel, point it at the claude/adrien-dashboard-juliano branch, and add the environment variables above.