ViralForge turns viral shorts into editable VFML projects
A TypeScript repo that gathered 301 stars in its first week deconstructs short video scripts into a structured report and generates ten costed variants.
What landed on GitHub
ViralForge, an agent-native viral video studio, was created on GitHub on 2026-11-11. By 2026-11-18 it had 301 stars and 9 forks, all gathered in those seven days, and it was appearing in GitHub trending AI repositories. It is written mainly in TypeScript and released under Apache-2.0. The repository's topics cover short video generation, video script deconstruction and AI agents.
The pitch, as set out in the project's README, is narrower than a general video generator. You give it a link or a local file. It returns a Viral DNA report that breaks the script into parts: the hook, a shot-by-shot structure, word anchors, a beat grid, caption style, an emotional curve and the conversion logic underneath. You then swap in your own product and script, and it rebuilds the original structure as an editable VFML project. From that project it generates 10 costed variants for A/B testing. The README is explicit that the output is meant to be an editable project, not an opaque MP4.
The repo lists three agent surfaces: a Skill, an MCP server and a CLI. FFmpeg appears among the project keywords. Version 0.1 ships with Chinese-language demo fixtures, while the documentation and agent surfaces are English-first, with Chinese documentation kept in a separate file.
The demos were made by hand
This is the part a builder should read before getting excited. The demo clips in the README were created manually with Jimeng Seedance 2.0 mini, using the original clip as a video reference. The README states plainly that they illustrate the intended workflow and are not an automated ViralForge model integration, and that exact shot timing has not yet been verified. The sample pairs are about ten seconds long, 16:9, at 720p.
The three worked case studies carry a similar warning. Their view counts, retention figures and DNA scores are fixture values that demonstrate the report format, not customer results. The per-variant costs are CNY demo estimates, and the README says the app does not bill.
The three case studies
The examples run from 13 to 15 seconds. GLOW LAB is a beauty serum film at 15 seconds with a DNA score of 92, a replica cost of ¥1.86 and a full set of ten variants totalling ¥4.34. AER is a tote bag film at 14 seconds, score 89, replica ¥1.72, variants ¥4.21. MORI is a rental makeover around a lamp at 13 seconds, score 94, replica ¥1.64, variants ¥4.10. Each dossier pairs a reference with a replica script broken into time ranges and shots, plus a table of ten hook and angle combinations aimed at different audiences. The hooks are written as lines of copy, for example "Stop paying the stupid tax" for the serum and "My landlord's beat-up rental? I fixed it for 30 yuan" for the lamp.
The README says you can run the studio locally with npm run dev and pick a case from the interface.
Limits and open questions
The README does not name the model or models behind the DNA extraction, so it is unclear whether the deconstruction runs locally, through an API, or partly by hand. It also does not explain what VFML contains beyond being an editable project format. The gap between the described one-click replication and the manually generated demos is the biggest open question: this suggests the generation half may still depend on a separate video model and a human in the loop. Whether the listed variant costs correspond to real generation spend, or are planning estimates only, is also not stated. What the source does confirm is the licence, the language, the star count and the feature list.
What it means for builders
For people shipping software, the packaging is the interesting part rather than the video output. Exposing a workflow as a Skill, an MCP server and a CLI at the same time is a pattern aimed at agents as the caller, and it likely points to more tools shipping an agent-callable surface alongside a normal UI. If you are wiring AI features into an app you are building, adding AI to your app covers that side of the work in Draftship, which builds and hosts apps from plain language with a built-in database, sign-in and hosting. Draftship is not affiliated with ViralForge and does not support it.
Primary sources
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.