andyhuo520/digital-human-koubo

This is an Agent Skill rather than a standalone app: it encodes a mouth-piece video production workflow so an agent can go from a script to a rendered MP4 with a consistent AI presenter. Try it if you already use a skills-capable agent and are willing to pay a generation service; skip it if you want a self-contained editor or have no appetite for credit-based generation.

The problem

A creator with a written script wants a finished talking-head video: a presenter who looks the same across uploads, lip-synced delivery, clean audio, captions, and the overlays that make a video feel produced. The usual path is stitching together a generative video tool, a separate editor, caption software and a packaging template, then redoing the same setup each time. Here the persona (face, outfit, voice description, prompt template) is stored in persona.json, and the agent takes the content through scripting, segment generation, QA, voice unification and a Remotion edit. The result is a finished MP4 rather than a folder of clips.

What's different

The distinguishing mechanism is that the project treats video production as an agent-runnable pipeline with explicit checkpoints. The script is split into segments of 15 seconds or less, generated through Renoise MCP (Grok Video 1.5 by default, with Wan 3.0 and OmniHuman documented), downloaded with CDN-safe range requests, back-checked with whisper, and frame-checked before editing. The agent quotes credits and waits for approval, then generates only the first segment before the rest, which gives a natural place to catch a bad clip before spending on the batch. This is more about gating costly steps than about a novel renderer.

The cost of that structure is dependency on a specific paid service and a long local toolchain. Renoise is the generation path and billing runs through the user's own account, so the workflow is not model-portable in the way a pure open-source renderer would be. The five editing styles and 48 packaging components are opinionated presets; they shorten the route to a packaged video, but they also define the look you start from. Remotion handles rendering under its own license, which matters to larger teams.

Who should try it

A good fit is a solo creator or small team already producing short explainer or ad-style mouth-piece videos, who has Claude Code or Codex installed, is comfortable giving an agent shell access, and would rather approve a credit quote than operate a timeline editor. The stored persona matters most for people who publish repeatedly with the same on-camera identity. A poor fit is someone who wants a self-contained desktop editor, needs a fully open generation stack, or works in an environment where agents cannot load skills or connect to a remote MCP server.

The supported-agent table is the real filter. Claude Code is listed as tested and the repo was built with it; Codex is expected to work with the same scripts; OpenClaw and WorkBuddy are marked as should work. Claude desktop or claude.ai can handle script writing, scene design, generation and quoting but only in a sandbox, so local editing needs a shell-capable agent. The README also documents Doubao, Qwen Office and NamiWork as options only if they support all three requirements.

Requirements and limitations

Running the whole pipeline needs three things together: an agent that loads Agent Skills, a connection to the remote Renoise MCP server (browser login, no API key), and local command execution with ffmpeg, Python and Node for downloading, QA and editing. Missing any one narrows what you can do, as the Claude desktop row shows. Generation consumes Renoise credits from your own account, and the workflow deliberately asks you to approve a quote, so budget is part of the process rather than an afterthought.

Ongoing upkeep includes local dependency versions, whisper-based checks and the Remotion render path. Remotion is free for individuals and companies of up to 3 people; larger companies need a company license. Animated emoji from Google Noto require credit in your videos, and maps have their own data and border-check notes. The third-party notices are the place to check before publishing branded or paid content.

HotRepo's take

The only dated signal is a NEW-REPOSITORY board observation on 2026-10-26: 356 total stars and #2 in the TypeScript new-project board, which is not a daily growth measurement. There is no comparable daily history, so persistence cannot yet be judged. Before committing, verify that your agent and environment can satisfy all three requirements, especially remote MCP plus local shell.

Project source and documentation