Price terms as published
Each line is the vendor’s own wording, split from the price capture checked on 2026-10-28. Nothing here is rounded, converted or re-worded — the full sentences stay in the review below. Lines that carry only a price continue the plan named on the line above, in the vendor’s own order.
| # | Plan / term | Checked |
|---|---|---|
| 1 | Free (MIT) | 2026-10-28 |
| 2 | generation billed in Renoise credits | 2026-10-28 |
One-liner: One photo, one sentence, one finished talking-head video — built as an Agent Skill that drives a paid video MCP for the lip-sync and Remotion for everything around it.
Evidence level: GitHub API plus the repository README read 2026-10-28. We did not run the pipeline; the cost and quality figures below are the project’s own and are labelled as such.
What it automates
The premise is narrow and useful: keep one consistent presenter. Your persona.json holds the face, outfit, voice description and prompt template, so every video stars the same person rather than a fresh face each time — composite the presenter into a studio, a shop or a street with an image model first, then make her talk.
Around that core, the repository automates the parts that normally eat an afternoon:
- Segmented generation. The script is split into segments of 15 seconds or less and generated through Grok Video 1.5 via Renoise MCP. Wan 3.0 and OmniHuman (for your own recorded footage) are documented alternatives.
- A cost gate before spending. The agent quotes segments, seconds, credits and remaining balance first and waits for a yes; it generates segment one for review before batching the rest.
- Quality control that is actually implemented. Each segment is transcribed with whisper and compared against the script, and four frames per segment are inspected — the README notes Grok sometimes burns subtitles into the picture, and those segments get re-shot.
- Voice unification. Spectral matching at third-octave resolution, de-essing, one shared small-room reverb and loudness normalisation, so thirty separately generated clips sound like one recording. The repository reports sibilance spread dropping from 13.2 dB to 5.9 dB in its own tests.
- Five editing styles and 48 packaging components. The components were measured frame-by-frame from 25 videos by creators including MrBeast, MKBHD, Kurzgesagt and Vox — a specific, checkable claim rather than “cinematic vibes”.
- Multi-language, tested in English, Japanese, Korean, French, Spanish and Arabic with whisper back-checks reported at 94–100%.
How to run it
The installation is a clone plus copying two skill folders into your agent, then letting the agent do setup checks for Renoise MCP, ffmpeg and whisper:
git clone https://github.com/andyhuo520/digital-human-koubo.git
cd digital-human-koubo
mkdir -p ~/.claude/skills && cp -R skills/digital-human-koubo skills/finance-explainer ~/.claude/skills/Codex uses ~/.codex/skills/ and codex mcp add for the MCP half. The repository’s own compatibility table is unusually candid about what is verified: Claude Code is tested (the project was built with it), Codex is expected to work on the same scripts, OpenClaw, WorkBuddy and Chinese agent products are “should work”, and Claude desktop / claude.ai is labelled sandbox-only — you can plan, quote and generate there, but the local editing steps need a shell. Three requirements define that line: the agent must load skills, must connect a remote MCP server, and must be able to run local commands.
The pipeline itself is eight steps long and the README documents each: diagnose content and recommend two of five styles, write script.json at roughly 5.4 Chinese characters per second, quote credits, generate and check segment one, batch, range-download from the CDN, whisper back-check, unify voice, build the word-level timeline, edit in Remotion, render and check mid-transition frames at −14 to −16 LUFS.
What it costs
The code is MIT — free to use and modify. The generation is not: lip-synced clips are produced through Renoise, which bills in credits. The repository gives a concrete number rather than a range: the 27-second livestream ad shown in its demos consumed 482 Renoise credits end to end. That is the number to extrapolate from, and it scales with segment count and length, which is exactly why the cost gate exists.
There is also a local-tooling cost that is easy to miss: ffmpeg, Python, Node and Remotion for the edit, and macOS Vision for the “text behind the presenter” effect (person segmentation → transparent video → three-layer composite) — that specific feature is macOS-bound. Every person in the demos is a synthesised digital human and is labelled as such; the male presenter is the author’s own digital twin.
How it compares with other digital-human options
The alternatives split into two classes, and the split matters more than any feature table. Hosted avatar tools are subscriptions in a browser: you upload a photo or pick an avatar, type a script, and get a video back — no shell, no local tooling, and no visibility into how the pieces fit. Digital Human Koubo is the opposite trade: you assemble the pipeline yourself, generate through a paid API in Renoise credits, edit locally in Remotion, and keep the QA code.
Inside the repository, the documented alternatives for the generation step are Wan 3.0 and OmniHuman — the latter aimed at making your own recorded footage talk rather than synthesising a presenter from a still. Those are model choices within the same pipeline, not competing applications, which is unusual and useful: the repo does not pretend its default model is the only option.
The honest summary is that this is not a cheaper HeyGen so much as a different shape of tool — more control, more assembly required, and a paid per-second cost that you can see before you commit.
What it will not do
Three limits worth knowing before you plan a channel around it. It is not a one-click web app — the full pipeline assumes an agent with shell access, and a chat-only product will stall at the editing stage. It cannot fix a bad generation: the QA loop catches subtitle burn-in and script drift by re-shooting segments, which costs more credits rather than fewer. And it does not remove the writing: the styles and packaging are opinionated templates, but the script, the persona and the judgement about what the video should say are still yours.
Verdict
If you make presenter-led content at volume and you already pay for an agent-based workflow, this repository is the most complete open implementation of that job we have found — particularly because the QA steps are real code and the cost is quoted before it is spent, which is rare in this category. If you want a hosted tool, or you cannot run local commands, wait: the value here is the pipeline, and the pipeline lives in your shell.