TAU-HOME.COM
LOADING

One-Token Video Router 'Treg' Unveiled as Reverse-Engineering Tool 'rea-video' Crosses 8,400 Stars

Superdesign has launched Treg's video routing catalog connecting five major video models at 0% markup via a single token. Meanwhile, JuneYaooo's open-source sho

tau · November 16, 2026

#Treg #rea-video #AI-Video #Agent-Tools #OpenSource

One-Token Video Router 'Treg' Unveiled as Reverse-Engineering Tool 'rea-video' Crosses 8,400 Stars

On November 13, 2026, Superdesign Dev Pty Ltd officially unveiled a dedicated video routing page for its agent tooling platform Treg, offering unified access to five major video foundation models through a single base URL and one authentication token. Concurrently, an open-source analysis tool named "rea-video"—built to split finished video into individual shots and reverse-engineer their underlying creation methods—surpassed 8,400 GitHub stars roughly two weeks after its initial release.

Conceptual diagram of Treg AI video router interface and rea-video shot breakdown pipeline

Image source: https://sloptvnews.com/treg-video-router-rea-video-shot-breakdowns

The twin developments highlight a maturing workflow around video generation pipelines: one lowering the friction of dispatching render jobs across diverse model APIs, while the other tears down finished media frame by frame to inspect how each scene was constructed. Treg founder Jason Zhou announced the video catalog on X on November 13, while the team's official blog indexes the update under November 11.

Five Video Models Under One Token with Zero Markup

The technical appeal of Treg's video catalog lies in consolidating disparate generative APIs behind a single, standard interface.

Historically, automating multi-model video pipelines meant managing separate authentication keys, rate limits, and billing dashboards for ByteDance's Seedance, Google's Veo, Alibaba's Wan, MiniMax's Hailuo, and Kuaishou's Kling. Treg abstracts those proprietary protocols into one base URL. Coding agents running inside developer environments like Claude Code or Codex can invoke video generation alongside companion tools, including image generators like Nano Banana Pro, GPT Image, Seedream, and FLUX, as well as licensed synthetic voices from Fish Audio.

Billing mechanics form a central pillar of the service:

  • At-Cost Metering: Catalog requests are billed directly from a prepaid balance at each provider's official public list rate with 0% markup.
  • Automatic Failure Releases: If an upstream generation job fails, the associated balance hold is cleared immediately without charging the account.
  • Free Allowance and BYOK: New accounts receive $1.00 in free testing credit, and requests routed through a team's own registered provider API keys incur zero routing surcharges.

Model-by-model pricing reveals distinct trade-offs. According to Treg's comparison table compiled from Higgsfield, fal, and Runway list prices dated November 11, rendering a 14-second Seedance 2.5 clip at 480p costs $0.1028 per second through Treg versus $0.2205 per second on fal. In a four-step reference workflow showcased by Treg—locking character and set stills via Nano Banana Pro, adding voice from Fish Audio, and rendering the 14-second Seedance sequence—the total run cost was $1.68, with $1.44 going directly to video generation.

Treg is not universally the cheapest option, however. On MiniMax H3 at 768p, Treg's catalog lists a rate of $0.08 per second, which sits above fal's $0.06 per second. In addition, prospective enterprise adopters should note that while Treg's repository is published under an Apache 2.0 foundation, it includes custom license clauses prohibiting commercial hosted-service redistribution, necessitating a legal check before self-hosting.

How rea-video Breaks Down and Classifies Video Shots

While Treg streamlines the synthesis side, JuneYaooo's open-source tool rea-video—first committed on October 29, 2026—approaches the media stack from the opposite direction. It takes an exported video file and dissects how each sequence was produced.

Built to run with Python 3.10 or newer and FFmpeg, rea-video processes media through dual visual and auditory pipelines:

  • Shot Segmentation and Audio Isolation: The video is sliced into granular cut boundaries, while its soundtrack is automatically separated into speech, background score, and sound effects.
  • Shot-Level Origin Classification: Each shot is evaluated and classified into distinct categories: live-action footage, AI-generated synthesis, digital avatar, 3D CGI, code-rendered motion graphics, or desktop screen capture.
  • Deconstructed Remake Blueprints: The tool outputs a detailed remake strategy alongside a structured VIDEO.md package that automated agents can read layer by layer.

According to evaluation data published in the project repository, rea-video achieves between 98.3% and 99.6% classification accuracy on pristine source files without requiring secondary vision model passes. On recompressed videos downloaded from social media platforms, that standalone accuracy drops to roughly 89%, at which point the documentation advises passing ambiguous segments to a vision-capable multimodal model for verification.

Connecting Generation and Reverse Engineering

In practice, Treg and rea-video form complementary bookends of modern AI video production.

When creators analyze high-performing reference material, rea-video provides structural insights into pacing, visual styles, and the real proportion of AI synthesis versus conventional post-production. Once creators understand the structural blueprint and preliminary cost profile through VIDEO.md, they can assemble their own licensed assets and hand off the generation workload to Treg's unified routing layer.

By eliminating manual API onboarding across multiple cloud vendors and offering precise visibility into how existing clips were constructed, the pairing enables independent creators and small teams to experiment with multi-model video pipelines with significantly reduced operational overhead.

Sources

Impact-Site-Verification: 9b3ebd10-6c45-44cf-a014-5b8c19b7b285