Agents365-ai/video-podcast-maker: Topic to 4K Video Through a Coding Agent

Who video-podcast-maker is actually for

The problem it addresses is assembly, not creativity. Going from a topic to a finished narrated video means research, a script, speech synthesis, subtitle timing, stock or generated visuals, background music and a 4K encode. Each of those is a separate tool with its own format, and the glue between them is where most solo creators stop. This project ships that glue as a skill: one pipeline in which a coding agent drives research, script generation, TTS, a Remotion composition and FFmpeg mixing, then renders at 3840x2160.

The intended user is someone who already works inside a coding agent such as Claude Code, OpenClaw, OpenCode, Codex or Pi, and who publishes to Bilibili, YouTube, Xiaohongshu, Douyin or WeChat Channels. Output languages are zh-CN and en-US. The README states that no coding is required and that the agent guides each step interactively, but that framing is slightly misleading: the setup steps are shell commands, and the project expects a Remotion project on disk. It is closer to a producer's toolchain wrapped in a conversational interface than to a consumer app.

How the pipeline moves from topic to rendered 4K

The README describes the workflow as research, narration script, TTS audio, Remotion composition, then a 4K render with BGM. The pivot point is a single file, podcast.txt. Every downstream artifact is derived from it: the TTS narration, the subtitles, section transitions, animation timing and the final cut. Sections are marked with [SECTION:xxx] blocks, and the human-facing note in the README insists each block opens with a hook and lands a transition rather than reading as a bullet-point dump.

TTS is synthesized in-house. The Edge backend is free and needs no key; Azure is the paid alternative. Both are local backends in the sense that no external component skill is required to reach them. Subtitles are Remotion-native, meaning SRT is rendered in React at 4K, with legacy FFmpeg burn-in still available.

Visuals come from a manifest-based asset engine. Each video carries a manifest recording license provenance, and producers are user files, assetseeker stock, imagencn AI stills, videogencn AI B-roll and Hyperframes overlays. The README is explicit that paid generation always asks first, which is a cost gate rather than a performance feature. There is also a design learning step that extracts style profiles from reference videos or images and applies them automatically when topics match, and a vertical shorts path that cuts 9:16 highlights from long-form sections.

Installing video-podcast-maker with the skills CLI

The README gives an npx command that installs the full skill globally. Pointing at the nested path installs only the full variant; dropping the /skills/video-podcast-maker suffix installs all three variants (full, -lite and -nano). Cloning the repository is the alternative.

npx skills add Agents365-ai/video-podcast-maker/skills/video-podcast-maker -g

After that you need Python 3.8 or newer, Node.js 18 or newer, FFmpeg, and a Remotion project. The README shows the macOS line and notes the Ubuntu equivalent in a comment.

brew install ffmpeg node python3
pip install -r skills/video-podcast-maker/requirements.txt
npx create-video@latest my-video-project
cd my-video-project && npm i

The one-time cost is worth planning around. The README states that a fresh Remotion project downloads about 2.2 GB of npm packages plus a roughly 90 MB Chrome headless shell, and recommends reusing an existing Remotion project with node_modules already installed. Lottie support is optional and needs @remotion/lottie plus lottie-web, installed per project only if you use LottieAnimation.

Configuration means setting TTS_BACKEND and the API keys for whichever backend you chose. Then you tell the agent to create a video podcast about your topic, and it runs the workflow. Previewing happens in Remotion Studio with npx remotion studio src/remotion/index.ts, and the README says the agent waits for an explicit render 4K confirmation before the final render. Under a skills CLI install, the paths shown in the repository root map to the agent's ${SKILL_DIR}.

The podcast.txt bottleneck the README admits

The most honest part of the documentation is a section addressed to the human rather than the agent. It argues that a weak script renders into 4K garbage and that no downstream polish rescues it. The prescribed process is to read the script mentally as the narrator, revise it at least three times across typos, filler and rhythm, audit numbers, proper nouns and English terms separately because that is where most TTS mispronunciations live, and estimate length at roughly 280 zh-CN characters per minute or 150 English words per minute.

That is a real limitation, and it is a design consequence rather than a bug. Because narration, subtitles, transitions and animation timing all derive from one file, errors in that file propagate everywhere at once. A mispronounced proper noun is not a small defect; it appears in the audio, in the rendered subtitles and in the pacing of the section it sits in. The stated acceptance test is whether any line makes you wince when read silently. If you are not willing to do that pass, the pipeline will faithfully produce a 4K version of a script you do not want to hear. Pronunciation control is handled through global and per-project phoneme dictionaries for Chinese polyphones, which is the escape hatch when a name keeps coming out wrong.

The other boundary is the toolchain itself. This is not a browser-based online video podcast maker. It needs Python, Node, FFmpeg and a Remotion project on the same machine as the agent, and the first run pays a multi-gigabyte package download. If you want to make one video on a laptop you do not control, this is the wrong tool.

video-podcast-maker compared with a hosted podcast video maker

The natural alternative is a hosted service in the class of Descript or Riverside, where you record or upload, edit by transcript and export. The difference in approach is where the work happens. A hosted editor assumes you already have footage or a recording session; its editing model is cutting and correcting existing media. This project assumes you have neither, and starts from a topic. It generates the narration, synthesizes the voice locally through Edge or Azure, assembles visuals from the asset manifest and renders through Remotion.

The trade-off runs the other way too. A hosted editor gives you a timeline you can scrub, frame-accurate trims and no local dependencies. Here, the revision loop is textual: you edit podcast.txt, preview in Remotion Studio, and re-render. That is faster for script-level changes and slower for anything visual that the composition does not parameterize. The project also targets distribution to Bilibili, Xiaohongshu, Douyin and WeChat Channels with per-platform publish info, which most Western hosted editors do not treat as first-class destinations. If your audience is on those platforms, the fit is closer than the feature lists suggest.

Licence, maintenance and what an upgrade costs

The repository is MIT licensed, which permits commercial use, modification and redistribution provided the copyright notice and permission notice are retained. That is the standard reading of the text; whether it fits a specific product is a question for your own counsel, not something the README answers.

Maintenance is visible in the release history. v5.2.0 and v5.2.1 landed on 2026-07-31 and 2026-08-01, and v4.2.1 on 2026-07-18 with a changelog note about polish scorecard visibility and Step 9 taste wiring. The last push to main was on 2026-09-14, two days before this writing, and the repository is not archived. The README itself carries a note that the project is still under active development and may not be fully mature, which is the maintainers' own framing rather than an outside judgement.

Upgrade cost is concentrated in the skill files and the Remotion side. The README points to CHANGELOG.md for version history, and the release notes are where breaking changes to the workflow steps would appear. Because the pipeline is a set of instructions a coding agent follows rather than a long-running service, an upgrade mostly means re-reading the skill and re-testing the steps you rely on. The heavy npm and Chrome download is per Remotion project, not per upgrade, so reusing one project across videos keeps that cost flat.

Editorial conclusion

Adopt it if you already run a Remotion project and a coding agent, and you accept that the narration script is your job, not the model's. Skip it if you want a hosted online video podcast maker with no local toolchain, or if nobody on the team will rewrite podcast.txt three times. Before committing, verify three things: that your agent supports SKILL.md, that TTS_BACKEND and its keys are set for the backend you picked, and that your existing Remotion project has node_modules installed so the one-time npm download is not repeated per video.

Frequently asked questions

Official sources

1. Agents365-ai/video-podcast-maker on GitHub
2. Issues
3. License: MIT
4. README
5. Releases