~/codigoagente
AI agents

How to make videos with AI using Claude Code or Codex on a regular PC (low-cost guide)

You can make videos with AI using Claude Code or Codex without an expensive graphics card: the agent writes the script, generates the voice, animates the scenes, adds transitions and captions, and renders the MP4. This short guide brings together the open source tools we've already reviewed on the blog, tells you which steps can run in the cloud with your PC switched off and which ones need your machine on, and what each path costs. It's written for a regular PC with 8 or 16 GB of RAM.

We checked it by making three videos in a Claude Code cloud environment with no GPU: an animated finance explainer with a growing chart and counters, an app demo with a browser, cursor and zoom over real screenshots, and a minimal test to measure timings. All of them have a Spanish voice-over, music, transitions and captions.

36 s compound interest explainer (in Spanish): script, data, chart, voice, music and captions made by Claude Code in the cloud with HyperFrames and Kokoro · codigoagente

What you need

PieceMinimumWhat for
A coding agentClaude Pro (20 USD/month, includes Claude Code) or ChatGPT Plus (20 USD/month, includes Codex)Writes the script and the video code, runs the commands
Node.js 22 or laterFreeHyperFrames, the render engine
FFmpegFreeCutting, audio mixing and export
Python 3.10 or laterFreeLocal voice (Kokoro) and captions (Whisper)
PC8 GB of RAM works; 16 GB is more comfortable. GPU optionalOnly for the steps that can't go to the cloud

You don't need to know video editing or animation code. You should be comfortable opening a terminal and reviewing the commands the agent proposes before approving them.

Three ways to do it: cloud, your PC or paid services

Which part of the work can go to the cloud and which needs your machine on · codigoagente

1. Everything in the cloud, PC off. Claude Code on the web (claude.ai/code) works in a remote container with Node, Python and FFmpeg. There the agent can write the script, generate a free synthetic voice with Kokoro, lay out the scenes in HTML with HyperFrames and render the MP4. It fits videos without footage of you: explainers, intros, product promos, tutorials built from text and diagrams. When it's done, the agent pushes the video to your repository or the folder you name, and you download it from your phone.

2. On your PC, switched on. Anything that starts from material that only exists on your machine: recording your screen or camera, cleaning up a recording, captioning what you said, editing by hand or cloning your voice for free. An 8-16 GB PC without a GPU handles all of it; what changes is how fast transcription and cloning run.

3. With paid services, from any machine. If you want a natural cloned voice without a GPU, AI-generated images or clips, or custom music, you plug in an API. The agent still orchestrates; you only pay for usage.

Which tool to use for each step

All of them are public projects we've already reviewed. The links go to the article with installation, license and limits.

StepToolLicenseWhere it runs
ScriptYour agent; [OpenMontage](/en/openmontage-ai-video-production-claude-code) if you want full production pipelinesAGPL-3.0 (OpenMontage)Cloud or PC
Free synthetic voiceKokoro via `npx hyperframes tts` ([HyperFrames](/en/hyperframes-html-to-video-ai-agents))Apache-2.0Cloud or PC, CPU
Clone your voice[VoiceStudio](/en/voicestudio-open-source-voice-cloning-elevenlabs-alternative) or [Fish Speech](/en/fish-speech-free-ai-voice-cloning-elevenlabs-alternative)AGPL-3.0 / Fish Audio CommunityPC (slow without GPU) or rented GPU
Screen recording[Recordly](/en/recordly-open-source-screen-recorder) (auto zoom and whisper.cpp captions)AGPL-3.0PC
Cut pauses and caption[editonny](/en/editonny-claude-code-video-editing-skill)MITPC
Hand editing with the agent's help[Concat](/en/concat-open-source-capcut-alternative) over MCPAGPL-3.0PC
Animation and transitionsHyperFrames, [Bang Motion](/en/bang-motion-motion-graphics-ai-agent) or [video-shotcraft](/en/video-shotcraft-product-promo-video-claude-code)Apache-2.0 / MIT / Apache-2.0 (uses Remotion)Cloud or PC
Render to MP4HyperFrames, Remotion or FFmpegApache-2.0 / Remotion / LGPL-GPLCloud or PC

Two license notes that matter if you'll sell the video or use it in ads: VoiceStudio's default voice model (OmniVoice) is non-commercial, and Remotion requires a company license above a certain size. Kokoro, HyperFrames, editonny and Bang Motion have no such catches.

Minimal test: how the agent works, step by step

To see the bare flow, we first asked Claude Code, in a cloud session, for a 19-second 1920×1080 video with four scenes, a female Spanish voice, transitions and captions. Here's the flow it followed, with the real commands; the two examples in the next section use the same method.

The 19 s minimal test: it looks like a slideshow, which is exactly what you need to get past · codigoagente

1. A script split into sentences

The agent writes the script as a list of sentences, one per scene. Each sentence becomes one audio file, and its length sets how long the scene lasts. For hard-to-pronounce names, write them phonetically in the voice text and keep the real name on screen.

2. Free voice with Kokoro

terminalsh
pip install kokoro-onnx soundfile
npx hyperframes@0.8.170 tts "Este vídeo lo hizo un agente de inteligencia artificial, con el computador apagado." -v ef_dora -o assets/audio/p1.wav
ffprobe -v error -show_entries format=duration -of csv=p=0 assets/audio/p1.wav

The first run downloads the voices (about 27 MB). Each sentence took seconds on CPU and produced 3.5 to 4.7 s of audio. ffprobe returns the exact duration to place the scene and its caption. For English, use one of Kokoro's English voices (npx hyperframes tts --list).

3. Scenes, transitions and captions in one index.html

HyperFrames turns HTML with timing attributes into video. The transitions are a colored panel that sweeps across the screen between scenes, animated with GSAP. The captions come from the script itself: since every sentence has its own audio and duration, the agent places the text at the same moment as the voice, and nothing needs transcribing.

index.htmlhtml
<div id="wipe1" class="wipe clip" data-start="4.9" data-duration="1" data-track-index="3"></div>
<div id="cap1" class="caption clip" data-start="0.5" data-duration="4.7" data-track-index="4">
  <p>Este vídeo lo hizo un agente de inteligencia artificial, con el computador apagado.</p>
</div>
<audio id="vo1" src="assets/audio/p1.wav" data-start="0.5" data-duration="4.67" data-track-index="10"></audio>

4. Check and render

terminalsh
npx hyperframes@0.8.170 check
npx hyperframes@0.8.170 render -o demo.mp4

check passed with 35 of 35 text elements at sufficient contrast after one fix (see common errors). The render took 39.9 seconds for 570 frames on a 4-core machine with 15 GB of RAM and no GPU, very close to a modest PC. The MP4 is 1.4 MB.

A prompt you can copy

snippettext
Make a 60-second 1920x1080 video with HyperFrames about <topic>.
1. Write the script as short sentences, one per scene, and show it to me before continuing.
2. Generate each sentence's voice with npx hyperframes tts and an English Kokoro voice.
3. Measure each audio file with ffprobe and place scene, voice and caption using those timings.
4. Use a short transition between scenes and my palette: <colors>.
5. Run npx hyperframes check, fix the errors and render to MP4.
6. Push the MP4 to <folder or repository>.

For Shorts or Reels, ask for --resolution portrait (1080×1920) when creating the project with npx hyperframes init. HyperFrames also ships ready-made agent workflows such as /faceless-explainer (a 60-90 s explainer from a text, no face on camera) and /product-launch-video (a promo from a website).

Videos that add something: animated explainer and app demo

A video of titles fading in and out looks templated. What makes people watch to the end is visuals that explain what the voice says: a chart that grows with real data, a cursor that clicks where the voice points, a zoom into the detail. HyperFrames makes that possible because each scene is HTML and JavaScript, which is exactly what a coding agent writes best.

Finance explainer for YouTube (36 s)

The agent wrote the script, computed the data (100 USD a month at 7% a year, compounded monthly, for 20 years: 24,000 contributed, 52,093 at the end) and drew it:

  • 240 blocks filling one by one while a counter climbs to 24,000.
  • A chart of 20 stacked bars (contributions in blue, interest in green) growing year by year, with the balance counting up to 52,093.
  • The interest part lights up when the voice says "28,000 comes from compound interest", then a bracket marks the last 10 years: +34,785 USD, two thirds of the total.
  • A closing takeaway and a note that it's an illustrative example, not financial advice.

The background music was generated by the agent itself with a Python script (chords and bass synthesized with numpy), so there are no licenses to check. Render: 1 min 20 s for 36.5 s of video on the same no-GPU machine.

App or website demo (30 s)

To show a product you don't need to record your screen: the agent builds a mock browser with real screenshots and animates it like a launch video. In our demo of the blog itself:

30 s demo made from three screenshots of the blog: 3D entrance, scroll, clicking cursor, zoom and labels · codigoagente
  • The browser enters with a 3D tilt, and the address bar changes on each screen.
  • The home page scrolls smoothly to the featured posts and guides.
  • The cursor moves to the search box, clicks (with a colored ripple) and the camera zooms into the filters.
  • On an article, it zooms into the "View the repository" button with a label explaining it.

Render: 1 min 8 s for 30.5 s. If your app has a web version, the same works for a SaaS, a store or a mobile app (with phone screenshots in a vertical frame). For web products you also have video-shotcraft, with 150 ready-made shots.

The prompts we used

snippettext
Make a 40 s 1920x1080 explainer with HyperFrames about <finance topic>.
Compute the data yourself and show it to me before animating. The visuals must
explain what the voice says: counters going up, a growing chart and the key
figure highlighted when the voice names it. Kokoro voice, captions from the
script, soft music generated with Python and a "not financial advice" note.
snippettext
Make a 30 s demo of my app with these screenshots: <paths>. Put them in a
browser with an address bar, bring it in with a 3D tilt, scroll, move a cursor
that clicks <button> and zoom with a label on <key feature>. Voice, captions,
music, and close with <URL>.

For an 8-10 minute YouTube video, the method is the same with more scenes chained together. At the pace we measured (a little over two seconds of render per second of video on 4 cores), expect about 20-25 minutes of rendering.

YouTube videos with your PC on

For a tutorial or a long video where you or your screen appear, the cheap flow is:

  1. Record with Recordly: automatic zoom on clicks and a smooth cursor, free and watermark-free.
  2. Clean up and caption with editonny from Claude Code or Codex: it removes pauses, breaths and mistakes, normalizes the voice to −14 LUFS and adds word-by-word captions with local Whisper. Before cutting, it shows you a settings card to approve.
  3. Add graphics: motion scenes from editonny's catalog, or ask the agent for titles and transitions with HyperFrames (/talking-head-recut overlays titles and data on your recording; /embedded-captions only adds captions).
  4. Touch up by hand in Concat, which the agent can also drive over MCP.

The cheapest and most natural voice for YouTube is still your own, recorded with a decent microphone. A synthetic voice fits explainers and promos; a cloned one fits when you can't or don't want to record every video.

On a PC without a GPU, Whisper transcription runs on CPU and takes noticeably longer than the video itself; let it run while you do something else.

Cloning your voice: free locally or fast if you pay

  • Free on your PC: VoiceStudio runs on CPU with lightweight engines, needs about 10 GB free on Windows, and is slow. Pick an engine with a commercial license (VoxCPM2, CosyVoice 3 or Kokoro, per their model cards), never OmniVoice if you'll sell the video.
  • Hourly GPU: Fish Speech Drama 3 needs a 48 GB GPU; an L40 on RunPod cost about 0.82 USD per hour. Commercial use is free if your yearly revenue is under 500,000 USD.
  • API without a GPU: fish.audio or ElevenLabs. Upload a clean sample of your voice and the agent generates the audio through the API.

Only clone your own voice, or someone's who has given you written permission.

Paid tools that raise the quality (and how Claude guides you)

ServiceWhat it addsReference price (Nov 2026)
ElevenLabsVery natural voices and instant cloningStarter 6 USD/month with instant cloning; Creator 22 USD/month with professional cloning
fish.audioCheap cloning and TTS via APIFree plan with 8,000 credits/month; Drama 3 at 50 USD per million characters
OpenMontage providersAI-generated images and clipsA 60 s short from 1.33 USD per the README
RunPodHourly GPU to clone or generate locallyL40 48 GB from ~0.82 USD/h

The HyperFrames CLI already supports ElevenLabs voices, editonny uses it for music and sound effects, and OpenMontage asks for your approval before spending. You don't have to set anything up blindly: ask your agent to walk you through it.

snippettext
I want to use ElevenLabs for the voice in my HyperFrames videos.
Walk me through it step by step: where I create the API key, where I store it
as ELEVENLABS_API_KEY in a .env file that isn't committed to git, and how
I test it with a short sentence. Don't ask me to paste the key in the chat.

What each path costs

PathMonthly spendFor whom
Cloud with Kokoro + HyperFrames20 USD (your Claude or ChatGPT plan)Explainers, intros, promos and Shorts without recording yourself
PC with your voice + Recordly + editonny20 USD (same plan) + electricityTutorials and YouTube videos where you speak
Either one + ElevenLabs Starter26 USDNatural cloned voice without a GPU
Production with AI images and clips (OpenMontage)20 USD + 1-5 USD per videoVideos with generated scenes

The cheapest way to start is the first one: with the subscription you already pay for coding, the cost per video is zero.

Common errors

  • Failed to load npm/gsap@3.14.2/dist/gsap.min.js: net::ERR_TUNNEL_CONNECTION_FAILED in hyperframes check: the render browser can't reach the CDN (common in cloud environments with restricted networking). Copy GSAP into the project (npm pack gsap@3.14.2) and reference it as a local file.
  • Transcription skipped: getaddrinfo ENOTFOUND huggingface.co: Whisper couldn't download its model because the network blocks Hugging Face. In the cloud, use captions taken from the script; on your PC, download the model once and it works offline.
  • check fails on contrast in the middle of a transition: the transition panel covers text with a light color. Use a dark color for the sweep or mark it with data-layout-allow-occlusion.
  • The video comes out silent: every <audio> needs an id.
  • An element meant to appear later shows from the first frame (it happened to us with the click ripples): a GSAP fromTo inside the timeline paints its start state immediately. Add immediateRender: false to that animation.
  • editonny on Windows: not validated by the author; it may need Developer Mode to create the skill links.

FAQ

Can the agent make the video with my PC off?

Yes, if the video doesn't start from your own footage. In Claude Code on the web, the agent generates voice, scenes, captions and the MP4 in the cloud. Recording, editing your footage or cloning your voice for free needs your PC on.

Does it work with 8 GB of RAM and no graphics card?

Yes. Kokoro, HyperFrames, editonny and Recordly run on CPU. What you notice without a GPU is the speed of Whisper and voice cloning.

Does it work the same with Codex?

The editonny and HyperFrames skills work with both Claude Code and Codex, and Bang Motion with any agent that reads its SKILL.md. The flow is the same; only the command that calls the skill changes.

Can these videos be monetized on YouTube?

The code licenses don't restrict the videos you produce. Check the license of the voice model, music and images you use, and add your own value: YouTube doesn't monetize repetitive mass-produced content.

Which is better: synthetic or cloned voice?

For a short explainer, Kokoro is enough and free. For a channel where the voice is your brand, record your own voice or clone it with a paid service.

Conclusion

With your agent subscription and open source tools you can make animated explainers with real data and product demos in the cloud without turning on your PC, and YouTube videos with your own voice on an 8-16 GB machine. Start with the HyperFrames and Kokoro test and, when you record your first tutorial, move on to editonny for cuts and captions.

Help us grow

codigoagente is free, with no paywalls. If this article helped you, any of these keeps the blog going.

// newsletter

One email a week with the repos worth your time

The week's new articles, tested against real code. No spam; unsubscribe in one click.

AI agents

guide · The complete ECC guide for Claude Code · #8

AgentShield, ECC's scanner, checks CLAUDE.md, settings.json, MCP, hooks and agents. We ran it on a deliberately vulnerable setup and on ECC's own install.
  • Security and code review
  • Claude Code
read article
AI agents

guide · The complete ECC guide for Claude Code · #7

We measured ECC's context footprint with /ecc:context-budget --audit (about 12,500 tokens on the minimal profile) and how to trim it: model, compaction and MCP.
  • Claude Code
  • Skills for your agents
read article
AI agents

guide · The complete ECC guide for Claude Code · #6

What ECC's 28 hooks do, how GateGuard stops the first edit, how to save and resume sessions and how continuous learning instincts work, with a real test.
  • Claude Code
  • Skills for your agents
read article