LaplaceYoung/dsh-directorx

DirectorX is a DSH-native video agent plugin built exclusively for DeepSeek Harness, designed to support end-to-end AI video production workflows from text brief to final delivery. It leverages DSH’s core capabilities including agent loop management, context tracking, user approval, and workflow orchestration, while adding all the specialized tools, director domain knowledge, production recipes, FFmpeg-powered media processing pipelines, and an infinite storyboard canvas that AI video production requires. It installs as a standard dsh-plugin into DSH, and does not implement a separate independent agent runtime.

This plugin serves three main groups: developers looking to build reusable AI video agents on DSH, content creators who need to turn text briefs into finished videos, and enterprise teams that require auditable approval workflows and cost control for video production. A standard workflow starts with a user-provided text brief, then DirectorX works with DSH to generate a script, storyboard, and character sheet that require user confirmation before assets are generated to the storyboard canvas. After generation, all editing, color grading, audio mixing, and quality checks run locally via FFmpeg, so you can re-render the timeline without re-generating all assets.

DirectorX is released under the open-source Apache 2.0 license, and requires Node.js 22.19 or newer and DSH version 0.1.1-rc.2 or higher to run correctly. It supports all major text-to-image, text-to-video, and vision models out of the box, including DeepSeek’s first-party multimodal models, OpenAI-compatible endpoints, and popular commercial video models like Sora 2, Runway, and Google Veo. Even without a valid API key, users can run a full mock workflow to test all pipelines, canvas functionality, and editing steps before adding API credentials for actual generation.