Bring your own keys.
VidNexus connects to the services you already use. Zero markup, swap providers anytime, your API keys live on your machine encrypted at rest.
Run any open-weight LLM locally on your GPU. Zero cloud cost per script, complete data privacy. The default and recommended LLM provider for VidNexus.
- ▸Default: qwen3.5:9b (6.6 GB download, runs on an 8 GB GPU)
- ▸Hardware-matched model recommendation based on your GPU
- ▸Upgrades for bigger GPUs: qwen3.5:27b, qwen3.5:35b (MoE)
- ▸Alternative voice: gemma4:12b
- ▸qwen3.5:4b for low-VRAM and CPU-only setups
- ▸A free Ollama account key turns on live web research, so scripts cite real sources
- ▸Any model in the Ollama library works - bring your own
GPT-4 class quality scripts via OpenAI's API. The best option if you don't want to run a local LLM, or if you want top-tier script quality.
- ▸Supported: gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini, gpt-3.5-turbo
- ▸Recommended: gpt-4o for documentary-style scripts
- ▸Recommended: gpt-4o-mini for high-volume listicle channels (cheaper)
- ▸JSON mode used for scene-by-scene structured output
- ▸Streaming support for live script preview in the editor
Point VidNexus at any OpenAI-compatible endpoint - your homelab inference server, a vLLM cluster, LM Studio on your other PC, OpenRouter, Together AI, Anyscale, Groq.
- ▸vLLM with Qwen-72B, Llama-70B, or DeepSeek-Coder
- ▸LM Studio running any GGUF model on a separate machine
- ▸OpenRouter: one key for hundreds of hosted models
- ▸Together AI: Llama-3.1-70B-Instruct-Turbo, Qwen-2.5-72B-Instruct
- ▸Groq: llama-3.1-70b-versatile (high-speed inference)
- ▸Anyscale endpoints
- ▸Any service implementing OpenAI's /v1/chat/completions spec
Powers the AI editor (Ctrl+K) when a Claude API key is set - natural-language timeline edits with undo snapshots. Without a key, the AI editor falls back to your local Ollama.
- ▸AI editor runs on Claude Sonnet 4.5
- ▸"Make scene 3 punchier", "swap this clip" - plain-English edits
- ▸Every AI edit is undoable via snapshots
- ▸Adding new motion graphics from the AI editor needs Claude
Best-in-class neural voice synthesis, per-scene. Voice set per channel, brand kit, or video. The default voice provider for VidNexus; ai33.pro works instead.
- ▸30+ premade voices in the standard library (Adam, Bella, Charlie, Daniel, etc.)
- ▸Voice cloning: upload 1-min sample → clone in 30 seconds (Eleven Voice Lab tier)
- ▸Multilingual v2 model - 32 languages, natural prosody
- ▸Per-channel default voice + speed in brand kits
- ▸Per-video voice + speed override in Video settings
- ▸Voice stability + similarity sliders per voice
- ▸Per-scene MP3 caching - only re-synthesize what changed
The voiceover through ai33.pro instead of ElevenLabs direct: the ElevenLabs voice library plus MiniMax, Fish Audio and Vbee voices, for a fraction of the per-character price. Add the key that you use; with only an ai33.pro key, VidNexus uses it automatically.
- ▸Pick the voice library (ElevenLabs, MiniMax, Fish Audio, Vbee...) and search it in the voice picker
- ▸ElevenLabs voices you already picked keep working through ai33.pro
- ▸With both keys set, Settings > Voiceover service chooses which one is used
Massive HD stock footage library - the primary stock source. Searched alongside Pixabay, Coverr and the public-domain archives, CLIP-scored against your narration.
- ▸Library: 3M+ videos, 200K+ in 1080p+
- ▸Most clips also available in 4K
- ▸Categories: nature, business, technology, lifestyle, abstract, history (b-roll)
- ▸Curated weekly editorial picks
- ▸Pexels license - free for commercial use, no attribution required
- ▸Search via VidNexus is CLIP-scored - matches narration intent, not just keywords
Alternative stock source - often has clips Pexels doesn't, especially in historical/illustrative content.
- ▸Library: 2M+ videos, 5M+ images
- ▸Strong on historical/illustrative b-roll (vintage maps, parchment, etc.)
- ▸Pixabay license - commercial use OK, no attribution
- ▸Same CLIP-scoring as Pexels in VidNexus
- ▸Cycled in parallel - if Pexels doesn't have a clip, Pixabay often does
Curated cinematic stock footage. Smaller library, higher cinematic quality.
- ▸Library: ~5,000 hand-curated cinematic clips
- ▸All clips 4K-source, downscaled to 1080p for delivery
- ▸Strongest categories: tech, architecture, abstract, drone shots
- ▸Pulled in parallel with Pexels + Pixabay - Coverr typically wins on visual polish
Public-domain films from the Prelinger Archives (ads, educational and industrial films), FedFlix (US government films) and the Internet Archive's stock footage collection. The best source for a video set in the past.
- ▸Only these three collections: the rest of archive.org is uploads whose public-domain marks can't be trusted
- ▸Long films: VidNexus scores the film's own frames and captions and picks the minute that fits the scene
- ▸Downloads only that part of the film, not the whole file
- ▸Credits listed in the video description
Openly licensed video from Wikimedia Commons - strong on historical, scientific and archival subjects.
- ▸Public domain, CC0 and CC BY clips; CC BY-SA only if you turn it on in Settings
- ▸Title, author and license kept per clip and listed in the video description
- ▸Same CLIP-scoring as every other source
NASA's launches, missions, Earth views and space imagery, public domain.
- ▸Only footage credited to NASA (third-party clips NASA shows are left out)
- ▸Long videos are cut to the part that fits the scene
- ▸NASA credited as the source in the video description
US military public affairs video: aircraft, ships, training, disaster relief. Public domain US government work.
- ▸Each user needs their own key (DVIDS' terms don't allow a shared one)
- ▸DVIDS' credit line ("U.S. Navy video by ...") listed in the video description
- ▸Same CLIP-scoring as every other source
Clips you downloaded from a paid library. Drop them in the My footage folder and the build picks the ones that fit each scene.
- ▸Any sub-folders; video and photos
- ▸Indexed with CLIP on your PC, the file names count too
- ▸Used where they are, never uploaded anywhere
Optional cinematic camera-move shots generated by Higgsfield (hosted) for scenes where stock doesn't cut it. The only AI video feature in VidNexus.
- ▸Named camera moves on a reference image: push-in, orbit, parallax, dolly, crane, tilt
- ▸5-8 second cinematic shots used as a scene's main visual
- ▸Hosted by Higgsfield - no local AI video generation
Real map tiles for the map-based motion graphics - routes, regions, flyovers - rendered with Remotion.
- ▸Used only by map motion-graphic types
- ▸Everything else in the 100+ motion-graphic library works without it
AI-painted YouTube thumbnails. 1-8 variants per run. The default thumbnail provider for VidNexus.
- ▸1-8 thumbnail variants per run - different compositions/styles
- ▸Prompt suggested from your script, edit any variant with a prompt
- ▸Trained for high-CTR YouTube thumbnails (faces, text overlay, contrast)
- ▸Editable hook text per variant before final selection
- ▸1280×720 output (YouTube native)
- ▸Reroll any variant individually without re-generating the rest
Built-in YouTube upload through your own Google Cloud OAuth app. Your credentials, your quota - VidNexus never holds your channel access.
- ▸Title, description, tags, category, language
- ▸Visibility (private by default) + scheduled publish
- ▸Playlist, custom thumbnail, made-for-kids flag
- ▸Multiple YouTube accounts, each linked to a channel
- ▸Publish history per video
Zero markup. Forever.
You pay each provider directly at their published rate. VidNexus never sits between your wallet and their meter.