Give it a niche. Get back a creator.

Own-feed trend research. research_niche_trends scrolls the account’s own IG Reels and TikTok For-You inside its Dolphin profile and runs Claude Vision over each screenshot for hooks, triggers, and styles — about $0.02 a run, no web search.

Music from followed accounts only. Audio is scraped from the accounts you actually follow — never the algorithmic feed of ads and suggestions — downloaded through the profile’s own cookies, capped at 4 reels per account.

Speech filtered out. A librosa harmonic/flatness classifier drops voiceover-only reels, so the pool is real music, not someone talking over a beat.

Beat-drop alignment. A multi-signal detector finds the drop; the editor aligns the cut to it only when confidence clears 0.30, otherwise the track plays from the top.

Hooks database. Every hook and its real reach/likes/saves are mirrored to Postgres, keyed on the hook text — the memory the scorer and recommender read from.

The multi-cut is enforced in Python. Each phase must carry at least two clips — fewer raises an error — so no beat ever lands on a single static shot.

Remotion receives pre-baked MP4s. It paints overlays, hook text, flashes and transitions only; the four compositions all render 1080×1920 at CRF 16.

One caption style, everywhere. A single Python source of truth burns the TikTok-meme look — heavy white fill, black stroke at 10% of the font — via local rendering; no paid subtitle service on the path.

Tier-list face-safe layout. The seated creator fills the frame while the tier table fills the lower band — a coded zone convention (overlays sit below the face), prompted seating, not runtime face detection.

Music drop-aligned in post. After the cut, FFmpeg aligns the track to the detected drop, then adds optional grain and shake — raw by default.

Real Chrome via Dolphin Anty. Each account owns a persistent anti-detect profile — real Chrome, US residential proxy, logged-in cookies that survive between sessions — driven by Playwright over CDP.

Four platform uploaders. Instagram, TikTok, Facebook and YouTube each drive the real web UI; IG and TikTok poll for a genuine post-confirmation. Platform scope is set per account.

Human-paced driving. The poster types the caption, waits, scrolls and clicks — no headless shortcuts on the money path, and the “open a new tab first” pattern never kills the session.

Mac ↔ server bridge. When Dolphin runs on your Mac and the API on the VPS, a claim-locked publish queue hands each render to local Dolphin and posts the permalink back — no double-posts.

Readiness checks & dry-runs. A login-state probe confirms the profile is signed in before posting; TikTok supports a dry-run that stops before the final Post.

Paid autonomy is opt-in. Every paid background job — the per-minute posting tick included — short-circuits until you set the flag, and the scheduler reports itself stopped.

The approval gate is the same code path. When an account is review-gated, a render is queued instead of posted; approving re-fires the identical render. The cron path posts without bypass, so the gate is live on the autonomous path for the transformation, talking-head and faceless engines.

Blitz review. Generated reels land in a swipe-style approve/reject inbox — you clear the day’s queue in a minute, or hand the gate off per account when you’re ready.

Spend caps, two enforcers. A per-account daily cap is checked twice — one in-memory, one durable in Postgres that refuses to spend past the limit — plus a monthly cost throttle.

Hooks are scored before posting. In every hook-led format, each variant is rated hook·0.40 + history·0.40 + timing·0.20; the best has to clear a per-account threshold or the batch re-rolls.

Virality outranks taste. The recommender learns from approvals and outcomes — when the operator’s preference and the numbers disagree, the numbers win, unless you hard-veto.

Analytics close the loop. The monitor agent pulls real per-post engagement every 30 minutes back into the hooks database, steering the next render toward what your audience actually rewarded.

Cross-account learning. Signal from one account’s wins is shared into the recipe for the next render — CreatorOS gets smarter across the whole roster, not one handle at a time.

An append-only, insert-only ledger. Every paid provider call is one immutable Postgres row — a database trigger rejects deletes — split into client wallets and provider pools. Profit is Σ charged − Σ cost, read straight off the table.

Per-format receipts. A single pricing source of truth itemizes each render as step receipts, from the Whiteboard format at $0.111 up to a UGC ad at $4.43 — every paid call debited the moment it happens.

Two spend gates. An empty wallet returns 402; a dry provider pool returns 503 — the render stops before it can spend money that isn’t there.

Top-ups are admin ledger entries. A top-up writes a transaction row and a commission (25%, clamped) sets the token rate — there is no card processor and no self-serve payment path; funding is deliberate and auditable.

Bring-your-own-keys routing. A full customer key set can pay its own providers and bypass the wallet on the render path, with telemetry still firing — partial keys fail closed to the wallet.