Rewrites the topic brief only (hooks, topics, facts) — spoken scripts stay Python templates. Default is Groq free-tier (qwen/qwen3.8-27b). Gemini, Groq, OpenRouter free, Mistral Experiment, and NVIDIA preview NIM are free-tier cloud (not local GPU). Hugging Face is credit-limited (~$0.10/mo then paid) — not free for reliable VO. Claude is disabled in the picker (paid / not free). Grok (xAI) API is backlog / not selectable (SuperGrok Lite chat ≠ API). Ollama is free-local but can freeze this machine. OpenAI is paid (needs credits; no free Chat Completions for new keys). Groq is console.groq.com, not xAI Grok. NVIDIA uses NVIDIA_API_KEY at integrate.api.nvidia.com. Hugging Face uses HF_TOKEN / HF_API_KEY (also HF_key) at router.huggingface.co. OpenRouter uses OPEN_ROUTER_KEY at openrouter.ai (free :free models / openrouter/free). Mistral uses MISTRAL_API_KEY at api.mistral.ai (Experiment rate-limited). This app never runs ollama pull.

Provider

Model

OpenAI key: checking…

Claude key: checking…

Gemini key: checking…

Groq key: checking…

Grok (xAI) key: checking…

Hugging Face token: checking…

OpenRouter key: checking…

Mistral key: checking…

NVIDIA key: checking…

Ollama: checking…

Ollama local lifecycle

Free-tier usage

Groq shows live remaining requests/day (RPD) per model from a tiny probe. Gemini, NVIDIA, Hugging Face, OpenRouter, and Mistral do not expose remaining quota via API — open their dashboards. Cached ~10 minutes (Refresh forces a new Groq probe).

Usage: not probed yet.

Refresh free-tier usage

Edge TTS is the volume default (no key). Optional cloud voices appear in the Create / Queue Voice dropdown when keys are set in .env. Keys are never shown here — only present/missing.

ElevenLabs KEY1: checking…

ElevenLabs KEY2: checking…

Free ElevenLabs keys only get premade voices in the API (Adam, Antoni, Brian, Daniel, Will, George, Callum, Sarah, Matilda, Lily, Jessica, Alice). Voice Library / community voices return paid_plan_required (402) — they are hidden from the dropdown. Paid plan required for library voices or commercial ToS.

Fish Audio KEY1: checking…

Fish Audio KEY2: checking…

Fish free model (s2.1-pro-free) is for A/B / fair-use lab. Commercial finals prefer ElevenLabs paid or Fish paid — confirm ToS before monetized publish.

Agnes generation (video + images)

One setting for Modes 1 / 2 / 2.5 / 2.6: Agnes video (T2V + I2V) and most Agnes images (logos, other T2I). Default stays 2.0. 2.5 Flash uses a different video API (720P, text/keyframe+first_frame) and agnes-image-2.5-flash for non-thumb T2I. YouTube prompt thumbnails always use agnes-image-2.0-flash (2.5 Flash invents Rubik/cube Icon stills). Jobs already queued keep the video model they were stamped with. After Save, hard-refresh the page.

Agnes fallback

Free only is the current proven mode. Paid keys stay in .env as PAYED_AGNES_KEY / PAYED_AGNES_KEY_2 / PAYED_AGNES_KEY_3 / PAYED_AGNES_KEY_4 — never shown here. Shorts use A+C+D; long form uses B only.

Paid key A (shorts): checking…

Paid key B (long): checking…

Paid key C (shorts): checking…

Paid key D (shorts): checking…

Paid quota estimate: checking…

Mode 5.01 image providers

Build-time thumbs/stills for Mode 5.01. Keys live in .env — never shown here.

Cloudflare Flux token: checking…

Cloudflare account id: checking…

Automatic Flux thumbs and stills: checking…

Pollination key: checking…

Agnes image key: checking…

Mode 2 Agnes plan

Mode 2 text→video defaults to paid (A+C+D shorts · B long; rotate by estimated UTC daily seconds within each pool). Separate from Mode 1 free/paid fallback.

Default Mode

Create prompt and the New button start on this Mode (1 / 2 / 2.5 / 3 / 3.5). You can still change it per Build. Mode 3.5 keeps the vector wizard on the form.

Default engine

New Queue jobs start on this engine. You can still change it per job. Local GPU as a default does not load LTX from this page.

Agnes mode: free-only.

Named channel profiles (logo, description, handles, YouTube OAuth, auto-publish). Create a prompt picks a channel when more than one exists. Jobs show the channel name. To add one: click Add channel (it clears the Branding form and the button becomes Create channel), type the new name (description/language/logo optional), then click Create channel (or Save settings). Nothing is written to the selected channel meanwhile.

Branding

Logo watermark on every render. The last few seconds show the channel name and platform @ handle (YouTube, Instagram, or TikTok depending on the export).

Channel name Channel description

Optional About text (stored with branding). Empty is fine for new channels.

Video language

Drives VO script, captions/hashtags, and Edge TTS locale for this channel.

Main voice

Pre-selected as the Voice on New prompt for this channel (you can still change it per Build). The choices follow the channel language, same list as the Voice dropdown. None keeps the page default.

Logo path Logo brief (optional)

Social platforms

Start empty on new channels. Add a network to configure handle, Short/Long, optional SQLite secrets (eye to reveal; empty falls back to .env), Connect / Mark notes, and auto-publish. Named Cloudflare tunnels keep a stable hostname; ./scripts/restart_with_tunnel.sh keeps redirect URIs in sync — paste Meta / TikTok once (or again if using a quick tunnel).

Checking…

No social platforms yet. Click + Add social.

Applies to every page, including How it works.