fendouai/ViralForge

What it does

ViralForge turns a viral short video into a Viral DNA report: hook, shot-by-shot structure, word anchors, beat grid, caption style, emotional curve and conversion logic. You then swap in your own product and script, and the project rebuilds the winning structure as an editable VFML project, generating 10 costed variants for A/B testing. It is aimed at growth teams, UGC agencies, e-commerce brands and solo creators who want a repeatable workflow rather than a prompt-to-MP4 black box.

Shots are anchored to words rather than seconds, so changing a line re-flows shots, captions and voice together. The studio is also exposed as an agent toolchain: a Skill, an MCP server and a CLI, with machine-readable artifacts (dna.json, project.vfml.json, variants CSV, cost ledger).

Quickstart

The web studio requires Node.js 22.15+ and npm:

git clone https://github.com/fendouai/ViralForge.git
cd ViralForge
npm install
npm run doctor      # optional: check local prerequisites
npm run dev         # http://127.0.0.1:5173

Open the studio, pick one of the three built-in cases (GLOW LAB, AER, MORI), and walk the Deconstruct → Replicate → 10 variants flow. The demo runs on pre-rendered, self-made assets; shot cuts, BPM and pacing are computed locally and are real, and no model or platform API key is needed.

To analyze your own local video without an API key, install FFmpeg, Python 3.9+, numpy and opencv-python-headless, then run:

node cli/viralforge.mjs analyze ./reference.mp4 --out dna.json
node cli/viralforge.mjs report dna.json

Agents can use the bundled Skill by copying skills/viralforge into .claude/skills/, or connect any MCP client to mcp/server.mjs via the stdio configuration in the README.

Constraints

Node.js 22.15+ is mandatory for the web app and build. FFmpeg/ffprobe, Python with OpenCV and Chromium/Playwright are optional and only needed for re-rendering, media checks and CLI analysis. v0.1.0 is an MVP: the demo's semantic insights are curated rather than live model output, replica films are pre-rendered, and batch variants currently produce scripts, covers and cost estimates. Model calls and batch rendering require a key or a worker. Platform link downloading is intentionally not implemented — only local uploads are supported. Face swap is excluded, reference footage never enters the output, and every output carries an AI-generated label. The open-source side omits category recipes, hosted models, character consistency and team billing, which the README attributes to a cloud offering. Licensed under Apache-2.0.

HotRepo's take

ViralForge entered the TypeScript new-project list on 2026-11-13 with 104 stars and 6 forks, gaining 4 stars over a roughly 1.5-hour window. That is a short observation and does not establish a trend. The more concrete signal is the scope of the v0.1.0 release: a working local studio, CLI, MCP server and Skill, three documented case studies with variant matrices, and a clean word-anchored project format. Read the case metrics as fixture values demonstrating the report format, not as customer results, and treat the curated semantic layer as a demo rather than a model capability.

Project source and documentation