Digital Human Koubo: AI talking-head videos from one photo with an agent skill
Digital Human Koubo is an Agent Skill that turns one photo and a short brief into a finished talking-head video with a digital human presenter, using Renoise MCP generation and Remotion editing.
What is Digital Human Koubo?
Digital Human Koubo, in the repository andyhuo520/digital-human-koubo, is an Agent Skill that lets an AI agent produce a finished talking-head video, with script, lip-synced presenter, voice clean-up, captions and packaging, from a single photo and a short brief.
The project keeps a fixed presenter, with face, outfit, voice description and prompt template, in persona.json, so every video shows the same person. The agent can composite that presenter into a scene first, then make the presenter talk. The script is split into segments of no more than 15 seconds each and generated with Grok Video 1.5 through Renoise MCP, and other models, including Wan 3.0 and OmniHuman for the reader's own recording, are documented in the README.
After generation, the skill checks each segment, unifies the voice across clips, builds a word-level timeline and edits the result in Remotion, with a cover spliced in as the first frame. The repository had about 2,000 GitHub stars as of November 18, 2026, and its demo videos were all produced with this repository, with every presenter shown as an AI digital human.
Who is Digital Human Koubo for?
Digital Human Koubo is for creators and developers who already use an AI agent that can load Agent Skills, connect a remote MCP server and run local commands, and who want talking-head videos from a photo and a written brief.
Claude Code is marked as tested in the README, Codex is expected to work because it uses the same scripts, OpenClaw should work, and WorkBuddy should work, with menu names that may differ by version. Claude desktop and claude.ai can do script writing, scene design, generation and quoting in a sandbox only, while local editing needs a shell agent. An agent that can only do MCP plus chat can still plan, quote and generate clips through Renoise, with editing done with Claude Code or Codex, or by hand with the commands in SKILL.md.
The local tools are listed for macOS, Windows and Linux, except that the text-behind-presenter effect needs macOS. English, Japanese, Korean, French, Spanish and Arabic are listed as tested, with a whisper back-check of 94 to 100 percent. The five styles cover finance and technology explainers, influencer talking-head videos, launch breakdowns, step-by-step tutorials, and multi-host and multi-language videos.
Is Digital Human Koubo free and open source?
Digital Human Koubo is open source under the MIT license for its code and documentation, but video generation spends Renoise credits on the reader's own Renoise account.
- Platform: macOS, Windows and Linux for local tools; the text-behind-presenter effect needs macOS.
- Minimum versions: Node.js 18 or later for Remotion, plus Python 3, ffmpeg, Playwright with Chromium, and whisper.cpp for local editing and checks.
- License: MIT for code and documentation; Remotion has its own license, free for individuals and companies of up to 3 people, with a company license needed for larger companies.
- Cost: generation is billed to the reader's Renoise credits, and the agent quotes segments, seconds, credits and balance first and waits for approval.
- Language: the repository is listed as TypeScript, and the skill combines a SKILL.md file with Python, shell and Remotion scripts; the README is in English with a Simplified Chinese version linked from it.
- Latest release: v1.0.0, published on October 26, 2026, the only release listed as of November 18, 2026.
What stands out
- A fixed presenter from one photo: the presenter details stay in persona.json, and the agent composites the presenter into a livestream studio, shop, classroom or Kyoto street before generating speech.
- Segmented generation with a cost gate: the agent quotes segments, seconds, credits and balance first, waits for approval, generates segment 1 for checking, and only then batches the rest.
- Segment checks and voice unification: every segment is transcribed with whisper and compared with the script, 4 frames per segment are inspected, segments with burned-in subtitles get re-shot, and voice unification reduced sibilance spread from 13.2 dB to 5.9 dB in the README's tests.
- Five editing styles with fixed ratios and lengths: Style A is 16:9 and 3 to 20 minutes, Style B is 9:16 and 30 to 120 seconds, Style C is 16:9 and 30 to 90 seconds, Style D is 16:9 and 3 to 8 minutes, and Style E is 16:9 and 3 to 7 minutes.
- 48 packaging components and a default cover: the components were measured frame by frame from 25 videos by top creators, each style has a default packaging recipe, and every video gets a cover by default, spliced in as the first 0.5 seconds.
How do I install Digital Human Koubo?
Digital Human Koubo installs by cloning the repository, copying its two skill folders into an agent's skills directory, connecting Renoise MCP, and checking local tools with the included script.
From the project's README at https://github.com/andyhuo520/digital-human-koubo/blob/main/README.md, with releases at https://github.com/andyhuo520/digital-human-koubo/releases, the commands below are in the README's order.
Get the code.
git clone https://github.com/andyhuo520/digital-human-koubo.git
cd digital-human-kouboFor Claude Code:
mkdir -p ~/.claude/skills && cp -R skills/digital-human-koubo skills/finance-explainer ~/.claude/skills/For Codex:
mkdir -p ~/.codex/skills && cp -R skills/digital-human-koubo skills/finance-explainer ~/.codex/skills/For WorkBuddy, OpenClaw or other agents:
cd skills && zip -r ../digital-human-skills.zip digital-human-koubo finance-explainer && cd ..For Claude Code:
/plugin marketplace add ArcoCodes/renoise-plugins-official/plugin install renoise@renoise-plugins-officialFor Codex:
codex mcp add renoise --url https://renoise.ai/api/public/mcpcodex mcp login renoiseFor OpenClaw:
openclaw plugins install renoise --marketplace ArcoCodes/renoise-plugins-officialCheck local tools:
bash ~/.claude/skills/digital-human-koubo/scripts/doctor.shSound effects:
python3 ~/.claude/skills/digital-human-koubo/scripts/make_assets.pyA Renoise account is created first, and login is in the browser, with no API key pasted. The latest official Renoise instructions are at https://renoise.ai/plugin, and the reader can check the connection by asking the agent to check the Renoise balance. The local tools table lists ffmpeg, Python 3 with numpy and Pillow, Playwright with Chromium for cover layout, Node.js 18 or later for Remotion, whisper.cpp, and a whisper model file placed in the reader's models folder.
The repository ships no third-party audio, and the sound generation step creates an original set of 30 sound effects and a placeholder background track from code. Xiaomi MiMo TTS is optional, for a designed voice, the reader's own cloned voice, or narration-only videos, and it uses the reader's own account at https://platform.xiaomimimo.com/.
How it was built with Claude
Commits in the repository carry a Co-Authored-By trailer naming Claude, and the README says the repository was built with Claude Code. The README's supported-agents table marks Claude Code as tested and says this repo was built with it.
Try it with your AI. Try it with your own AI assistant:
Help me use the Digital Human Koubo Agent Skill from andyhuo520/digital-human-koubo to
make a talking-head video about [my topic] for [my audience], in [language] and [ratio
and length]. Use only the official commands and interfaces named in the README,
including Renoise MCP, Grok Video 1.5 and Remotion, and the repository's doctor.sh setup
check. First show me the setup check results, the style recommendation, the script
segments of no more than 15 seconds each, and the quote for segments, seconds, credits
and balance, and wait for my confirmation before you generate anything. Use only [my own
photo or a photo I have permission to use] for the presenter, keep the AI digital human
label on the video, and do not impersonate a real person. Show risky steps first and
wait for me to confirm. Do not install anything from unknown sources, do not touch
credential files, and do not delete anything. If a step needs Node.js 18 or later,
ffmpeg, Python 3, Playwright or whisper, tell me what is missing before continuing.What are Digital Human Koubo's limits?
Digital Human Koubo's limits are that generation needs a Renoise account and credits, editing needs local tools, and presenters and voices may only come from material the reader owns or has permission to use.
- Credits are spent only after the reader approves the quote, and the material identifiers in this repository belong to the author's account and will not work on the reader's account.
- The text-behind-presenter effect uses Apple Vision person segmentation and needs macOS; everything else is cross-platform, and it works on medium shots with headroom, while a selfie close-up hides the text.
- Grok sometimes burns subtitles into landscape clips, whisper writes Chinese numbers as digits, and swapping in a different voice afterwards leaves about 100 ms of lip offset, which the README says is fine in picture-in-picture and visible in full-screen close-ups.
- Use only the reader's own face and voice or ones with explicit permission, never impersonate a real person or creator, keep the AI digital human label on every video, mark scene and animal clips as AI generated, and do not invent numbers, prices, offers or reviews.
- Demo videos and presenter images depict the author's own AI digital humans for demonstration only, so the reader must not re-upload them as the reader's own work or use the presenters' likeness in the reader's own content.
- Animated emoji need the credit Animated emoji: Google Noto in videos that use them, maps involving China's borders in formally published content must be checked against applicable local standards, and voices may only be cloned when the reader owns the voice or is licensed to use it.
Related on claudefox
- Agent Autoloop: run Claude Code and Codex in the background and get a card when they need you
- Kanzo puts a kanban board and an MCP server for coding agents in one 4.7 MB binary
- External sandboxes for coding agents: run Claude Code, Codex and Pi commands in a Linux guest with npx or uvx
- ex-USSR Kinship Research: an evidence-first genealogy skill for Claude Code and Codex
Problems and feedback
Report problems and feedback on the repository's Issues page at https://github.com/andyhuo520/digital-human-koubo/issues.
Source: andyhuo520/digital-human-koubo repository, README, Releases, Issues, SKILL.md, Install reference, THIRD_PARTY.md, NOTICE.md, LICENSE, Renoise plugin instructions, Xiaomi MiMo platform
This story first appeared in the claudefox daily briefing for November 18, 2026.