Video Forge

Video Forge is an open-source toolkit from GitHub user Sanidhya14321 for producing editable video from a brief or a recording. A compact directing skill coordinates planning, motion, editing, captions, audio, generation and delivery, backed by local Remotion/FFmpeg execution. It works with Claude Code, Codex, Claude custom skills, ChatGPT and other compatible local agents. The latest release listed is v0.5.1.

What It Is

Video Forge is a set of agent skills plus a local rendering runtime. You describe a video, for example a 30-second launch reel, and the skill picks a visual language and the relevant modules, then renders through the local toolchain. Without an executable environment it delivers the production plan and job rather than labeling an unrendered plan as an MP4.

Workflows and Outputs

The README documents three production workflows: screenshot product demos, supplied-state algorithm explainers, and captioned footage repurposing, plus native short-form reels. It offers twelve motion families in two treatments, abstract minimal or colorful comic/character choreography. Editing supports explicit segment edits and optional threshold-based silence removal, which does not do semantic footage selection. Audio and captions use local voice and transcription, music mixing, measured word highlighting and SRT files. Delivery includes project and renderer snapshots, staged assets and credits, transcripts, captions, thumbnails and MP4 presets with decoding checks.

Hosts and Setup

Codex and Claude Code install as plugins from a marketplace, while Claude chat uses an uploaded custom-skill archive and ChatGPT chat gets a companion for planning and a local render handoff. A standalone archive serves other Agent Skills hosts. Local use needs Node 22+, npm, and first-use network access for pinned dependencies and the browser. The CLI offers doctor, review, preview, run and inspect commands.

Providers and Licensing

Local Kokoro and faster-whisper are supported, along with opt-in budget-gated OpenAI speech and fal adapters. Local production needs no account or credentials. Code is MIT-licensed, while Remotion, model weights and assets retain their own terms.