Aginx

Context

Personal branding and content marketing hit the same wall: the idea is there, but shooting, editing, voiceover and subtitles eat hours. Aginx removes the manual labor from that chain — a short free-form brief goes in, a finished vertical 9:16 video for Reels, Shorts or TikTok comes out.

How it works

A four-stage loop, each an isolated step:

• Orchestrator — Claude takes the brief and returns a strict JSON script: timed scenes, visual prompts (English, cinematic), narration (Russian), mood, camera motion, a music prompt and an opening hook.
• Visual — by content type: broll (per-scene video), slideshow (images with motion) or talking_head (an animated avatar speaking the narration).
• Audio — voice synthesis and background music; narration is generated before the paid video step, so no budget is burned if TTS fails.
• Assembly — ffmpeg stitches the scenes, mixes voice and music, burns in ASS subtitles and outputs a single .mp4.

Engineering frame

• One source of truth — the script. The LLM returns typed JSON against a Pydantic schema; everything downstream reads structure, not free text. Less ambiguity, easier to debug.
• Providers with fallback. MediaProvider calls fal.ai first and switches to Replicate on error; voice via ElevenLabs / MiniMax. Models are a config parameter, not a hard-wired dependency.
• Idempotency and resume. The script and intermediate files are written to disk immediately; an interrupted job can be resumed — finished scenes are skipped, not recomputed.
• Order for the sake of cost. Cheap and risky steps run before expensive ones: script and voice first, paid video generation last.

Modes and control

• Three content types: broll, talking_head, slideshow — the orchestrator picks one or you pin it.
• Typer CLI: brief as a string or JSON file, fine voice tuning (stability, speed, pitch, emotion), --dry-run to preview the script only, resume to finish a job.
• Config, not code: models, resolution, fps, music volume and voice parameters live in config.yaml.

What I took away as an engineer

• The "LLM director + typed script + generative models" combo turns video production into a controllable pipeline rather than a chain of manual steps.
• The hard part isn't generation but orchestration: stage ordering, fallbacks, idempotency and cost control matter more than the choice of any single model.
• Aginx is the base for a content factory: the same scheme can be adapted to someone else's brand or product.