digital-human-koubo
Produces talking-head videos with a fixed AI digital-human presenter. It recommends one of five editing styles, writes the script, splits it into ≤15 s segments, and generates them through Renoise Grok Video only after quoting credits and getting user approval. It then downloads and QA-checks the clips with whisper, edits them in Remotion with 48 packaging components, and makes a cover frame.
What it can touch
| files | write, including EXA/TAVILY/SERPER/FIRECRAWL/UNSPLASH/PEXELS/PIXABAY keys in ~/.config/xiaolin-explainer/keys.env, ~/.config/ai-teardown/mimo.env, ~/.config/xiaolin-explainer/keys.env, ~/.claude/skills/, ~/.codex/skills/, ~/.cache/whisper/ |
|---|---|
| network | api.exa.ai, api.firecrawl.dev, api.pexels.com, api.tavily.com, api.unsplash.com, api.xiaomimimo.com, commons.wikimedia.org, github.com, google.serper.dev, huggingface.co, pixabay.com, registry.npmjs.org, registry.npmmirror.com, renoise.ai |
| shell | runs commands |
| secrets | MIMO_API_KEY |
What it actually does
The instructions tell the agent to: call the Renoise MCP (get_account, estimate_cost, create_tasks, download_results); run local Python, bash and Node commands (doctor.sh, build_tasks.py, rangedl.py, assemble.py, align.py, cutout_video.py, make_cover.py, lin_* scripts, npx remotion render, ffmpeg); download generated clips from the Renoise CDN; transcribe with whisper.cpp; and write project files (script.json, brief.json, tl.json, edit.json, renders, covers). Optional paths read API keys from ~/.config/ai-teardown/mimo.env (Xiaomi MiMo TTS and voice clone) and ~/.config/xiaolin-explainer/keys.env (Exa, Tavily, Serper, Firecrawl, Unsplash, Pexels search and stock keys). Setup may clone node_modules from other local Remotion projects and may switch the npm registry to a Chinese mirror. The docs contain no hidden-character payloads, no instructions aimed at the reviewer, and no persistence or exfiltration. Credit spending is explicitly gated on user consent. Most executable scripts were omitted from the bundle, so their exact behavior could not be verified; only doctor.sh was visible, and it only checks for tools and prints install hints.
What the check found
mediumChanges the agent's instructions or settings
Editing CLAUDE.md, AGENTS.md, memory files, hooks or MCP settings lets a tool keep steering the agent even after it's removed.
README.md:94 | **Codex** (CLI / desktop) | `~/.codex/skills/` | `codex mcp add` | ✅ | ✅ Expected to work (same scripts) |
mediumChanges the agent's instructions or settings
Editing CLAUDE.md, AGENTS.md, memory files, hooks or MCP settings lets a tool keep steering the agent even after it's removed.
README.zh-CN.md:94 | **Codex**(CLI / 桌面) | `~/.codex/skills/` | `codex mcp add` | ✅ | ✅ 预期可用(同一套脚本) |
mediumDownloads programs or archives while running
Files fetched at run time weren't part of this check, and they can change at any time.
skills/digital-human-koubo/scripts/doctor.sh:30 [ -n "$M" ] && ok "whisper 模型 $(basename "$M")" || need "whisper 模型 ggml-medium-q5_0.bin" "逐段回听、逐字对齐" "下载到 ~/models/:curl -L -o ~/models/ggml-medium-q5_0.bin ht
mediumSuggests switching the global npm registry to a third-party mirror
When npm cannot connect, the agent is told to run `npm config set registry https://registry.npmmirror.com`. This persistently changes the user's global npm configuration so that all future package installs come from a third-party mirror. The mirror is legitimate and widely used in China, but the change outlives this tool and should be made only with explicit user consent.
skills/digital-human-koubo/references/install.md:41 Remotion 不用单独装:`scripts/new_edit.sh` 起工程时会先从本机已有的 Remotion 工程克隆,没有才 `npm install`;npm 源连不上时换镜像 `npm config set registry https://registry.npmmirror.com`。
mediumCopies node_modules from arbitrary existing Remotion projects on the machine
The setup scripts are told to clone node_modules from whatever Remotion project they find locally before falling back to npm install. That pulls in unverified, possibly stale or tampered dependencies from other projects, and it means the scripts look through the user's directories for such projects. The scripts doing this (new_edit.sh, setup_engine.sh) were not included in the bundle, so how far they search could not be checked.
skills/finance-explainer/SKILL.md:23 - 引擎:`bash scripts/setup_engine.sh <项目目录>`(复制 templates/engine,node_modules 优先从本机已有 Remotion 项目 `cp -c` 克隆,失败再 npm install)。
skills/digital-human-koubo/references/edit-styles.md:3 几个模板都是 Remotion 工程,`node_modules` 用 `cp -cR` 从任一现有工程克隆(npm 源常 SSL 失败)。时间轴全部来自逐字对齐(tl.json / timeline.json),改文案后重跑对齐即可。
skills/digital-human-koubo/references/install.md:41 Remotion 不用单独装:`scripts/new_edit.sh` 起工程时会先从本机已有的 Remotion 工程克隆,没有才 `npm install`;npm 源连不上时换镜像 `npm config set registry https://registry.npmmirror.com`。
mediumLarge amount of locally executed code not available for review
The workflow runs many local scripts automatically: doctor.sh, new_edit.sh (which also auto-runs make_assets.py), build_tasks, rangedl, assemble, align, cutout_video, patch_words, lin_mix, make_cover, and the finance-explainer research.py and fetch_stock.py. A Swift helper is compiled on first use. Of these, only doctor.sh was included in the bundle. The others could hide network calls or reads of the key files, and they need a human look before listing.
skills/digital-human-koubo/SKILL.md:25 2. **本机环境**(要剪辑时才需要):`bash scripts/doctor.sh`。有的打 ✓ 不提;缺的它会按 macOS / Windows / Linux 列出多种装法。系统软件让用户自己装,Python 包可以在用户同意后帮装。
README.md:168 `new_edit.sh` also runs this automatically for every new project. For a better soundtrack, drop any track you have the rights to at `<project>/public/music/bgm.
THIRD_PARTY.md:12 | Apple Vision | macOS system framework | Apple SDK | Person segmentation for the text-behind effect (`scripts/personmask.swift`, compiled locally on first use;
Keep an eye on it
This check is of one commit. Watch digital-human-koubo and we'll email you if a later commit changes what it can touch.
Made this? Add the badge
[](https://plugsday.com/tools/andyhuo520-digital-human-koubo)About this check
Checked November 2, 2026 at commit f661a6c047bb, 917 files. Rules for known red flags, then two Claude reviews, one of them told to assume the author is malicious. A later commit hasn't been checked. How checks work.