How to Make Videos with Claude Code for FREE (2026 Guide)

TL;DR: Install Claude Code, Node.js 22 and FFmpeg. Run npx hyperframes skills update, then npx hyperframes init my-video, start Claude Code in that folder and describe the clip you want. Preview with npx hyperframes preview, render with npx hyperframes render --output final.mp4. To edit footage you already have, use npx hyperframes init my-edit --video talk.mp4 and ask Claude Code for a cut list first.

HyperFrames also offers cloud rendering on HeyGen, paid per credit. Skip it. The local render is the free part.

What you need

Three things:

• Claude Code. It needs a Pro, Max, Ultra, Team, Enterprise or Console account; the free claude.ai plan does not include it. Or use the free Ollama route.
• Node.js 22 or newer, and npm.
• FFmpeg, which turns the captured frames into an MP4.

Install Claude Code with the official installer:

# macOS / Linux
curl -fsSL https://claude.ai/install.sh | bash

# Windows PowerShell
irm https://claude.ai/install.ps1 | iex

Step 1: Install FFmpeg and check your setup

Install FFmpeg, then run npx hyperframes doctor; it checks your setup and tells you what is missing.

On a Mac:

brew install ffmpeg

On Windows, get a build from the FFmpeg download page. Its Windows section links to the builds (gyan.dev and BtbN). On Linux, apt install ffmpeg.

Then check everything:

npx hyperframes doctor

Here is part of the real output on my Mac. A cross is a hint, not always a blocker: Kokoro and Docker are optional.

  ✓ Node.js          v22.12.0 (darwin arm64)
  ✓ whisper-cpp      /opt/homebrew/bin/whisper-cli
  ✗ TTS (Kokoro)     Not installed (optional — local voice fallback)
                     pip install kokoro-onnx soundfile
  ✓ FFmpeg           ffmpeg 8.1.2 at /opt/homebrew/bin/ffmpeg
  ✓ Chrome           cache: ~/.cache/hyperframes/chrome/...

Step 2: Install the HyperFrames skills

Run npx hyperframes skills update so Claude Code knows how to write a video, then confirm with npx hyperframes skills check.

npx hyperframes skills update
npx hyperframes skills check

The check prints where the skills live and whether they are current:

  Location  ~/.claude/skills (claude-code)
 
  ✓ 4 current   ◦ 5 available on demand

Step 3: Create a video from a prompt

Create a project with npx hyperframes init, start Claude Code inside it, and describe your clip using the /hyperframes skill.

npx hyperframes init my-video
cd my-video
claude

Restart Claude Code (a new session) after installing the skills so it loads them. Then paste this prompt. It is the exact one that made the clip at the start of the video:

Using /hyperframes, make a 5-second vertical 1080x1920 clip: dark background, a terminal that types "one prompt, $0", amber accent. Then render it to clip.mp4.

On my run (Opus, paid plan) that was one prompt and one turn: 2 minutes 13 seconds from prompt to clip.mp4, a 5-second 1080x1920 video. Claude Code even found and fixed a cursor bug itself inside that turn.

How it works

Claude Code reads the skill, then writes the clip as HTML. On my run it wrote two files, index.html and compositions/terminal.html, animated with GSAP. Every frame of your video is a web page. HyperFrames opens that page in a headless browser (Puppeteer with a bundled Chromium), captures each frame, and FFmpeg encodes the MP4.

Step 4: Preview, change, render

Preview in your browser, ask for changes in plain words, and render the final MP4 on your own computer.

npx hyperframes preview

The studio opens at http://localhost:3002, plays the clip back, and reloads when the code changes.

Don't like something? Just say so. These are the two follow-ups I sent:

Make the text bigger.

Slow the typing down.

Claude Code edits the code, and you look again. This loop is the whole job.

When it looks right, render:

npx hyperframes render --output final.mp4

Real output from the 5-second clip:

◇  .../my-video/final.mp4
   509.9 KB · 5.0s video · rendered in 3.8s

For a quick test render, add -q draft. The quality options are draft, looks and delivery.

Want a longer video? Use the same loop, one scene at a time: ask for a title, then a diagram, then a checklist.

Step 5: Add a free voice (Kokoro) and word timings

Install Kokoro once with pip, turn a text file into a WAV with npx hyperframes tts, then get word timings with npx hyperframes transcribe.

Kokoro (Kokoro-82M) is a local model: no API key, and nothing leaves your machine. HyperFrames does not ship it, so install it first:

pip install kokoro-onnx soundfile

Without it, tts fails with The kokoro-onnx package is not installed. The error also says you can point HYPERFRAMES_PYTHON at a venv Python that has the packages. On my Mac it worked with Python 3.12; Python 3.15 could not resolve the dependencies.

Write your script in script.txt, then:

npx hyperframes tts script.txt --voice am_michael -o voice.wav

Other voices include af_heart (the default), am_adam, bf_emma and bm_george.

Now get every word with its timestamp. Transcription runs locally with Whisper (model small.en by default):

npx hyperframes transcribe voice.wav

The first words of the real transcript.json:

[
  { "text": "One", "start": 0.05, "end": 0.24, "id": "w0" },
  { "text": "prompt", "start": 0.24, "end": 0.71, "id": "w1" }
]

Claude Code uses those timings to land each animation on the right word, and to add captions.

Step 6: Edit your own footage, cut list first

Pass your video into a new project with --video, then ask Claude Code for a cut list and approve it before anything is cut.

A screen recording or a talking-head clip both work:

npx hyperframes init my-edit --video talk.mp4

HyperFrames copies the video in and transcribes the speech with Whisper on your machine:

Video: 1920x1080, 41.0s
Transcribing...
Transcribed: 134 words (41.0s)

It saves every word with its timestamp in transcript.json. The CLI docs say it also patches captions into the composition, but on version 0.8.226 it did not, so ask for captions yourself (Step 7).

Start Claude Code in my-edit and send this prompt:

Read the transcript. Find every pause longer than 0.5 s and every sentence I said twice. Show me the cut list and wait for my OK.

Always ask for the list first. On my run Claude Code did not trust the Whisper gaps, so it measured the pauses in the audio with FFmpeg's silencedetect filter:

ffmpeg -hide_banner -nostats -i talk.mp4 -vn -af silencedetect=noise=-35dB:d=0.5 -f null - 2>&1 | grep silence_

It found 12 pauses, reported "Sentences said twice: None found", and waited. Read the list, then say yes, or change it. My reply:

Yes, cut all 12 as listed. Add a tiny audio fade at every cut, so there is no click. Save it as talk-cut.mp4.

FFmpeg trims each piece you keep (trim and atrim), fades the audio at every cut (afade) so you don't hear a click, and joins the pieces back together (concat). The clip went from 41.03 seconds to 35.40 seconds.

Add a title card and a lower third

A second prompt worth stealing:

Add a title card at the start and a lower third with my name. Keep the footage untouched underneath.

Claude Code builds those as separate layers over your footage, so the footage plays untouched underneath. Then render it with npx hyperframes render.

Step 7: Add captions, and avoid the FFmpeg subtitles trap

Either render the captions through HyperFrames, styled however you like, or export a caption file with npx hyperframes transcribe --to srt.

Option 1: ask Claude Code to add captions to the composition from the word timings, and render. They are burned into the video.

Option 2: export a caption sidecar file (srt or vtt):

npx hyperframes transcribe talk.mp4 --to srt -o captions.srt

◇  Exported 109 words to SRT → .../captions.srt

ffmpeg -i talk.mp4 -vf subtitles=filename=captions.srt out.mp4

No such filter: 'subtitles'
Error opening output files: Filter not found

What Claude Code can't do

It can't judge taste. It does not know which joke landed, or which take felt right, so you still watch the cut. What it saves you is the boring part: finding pauses and typing timestamps.

The free route, honestly

If you skip Pro and run Claude Code on a local model in Ollama (setup in the free Windows guide, ollama launch claude needs Ollama v0.15 or newer), know three things:

• Context. The HyperFrames skills are long documents. Ollama recommends a 64k or bigger context window for Claude Code, and with under 24 GiB of VRAM the default is only 4k. Raise it, for example with OLLAMA_CONTEXT_LENGTH=65536.
• Speed. Run ollama ps while it works. If the PROCESSOR column does not say 100% GPU, it will crawl.
• Quality. Ask for one scene at a time, and check the output.

I made this video on a paid plan. Then I tried the opening prompt once on a small local model, qwen3.5:9b, with a 65536 context at 100% GPU:

NAME          SIZE      PROCESSOR    CONTEXT
qwen3.5:9b    9.8 GB    100% GPU     65536

It took 22 minutes 50 seconds, said "Done", and got it wrong: a landscape 1920x1080 video with static white text, no typing, no "$0" and no amber. That is one model and one run, not a verdict on every local model. So start small, and check what you get.

Checklist

• Claude Code installed (Pro or higher, or the Ollama route)
• Node.js 22 or newer
• FFmpeg: brew install ffmpeg on a Mac, a build from ffmpeg.org on Windows
• npx hyperframes doctor
• npx hyperframes skills update, then npx hyperframes skills check
• npx hyperframes init my-video, cd my-video, claude
• Prompt with /hyperframes
• npx hyperframes preview, ask for changes
• npx hyperframes render --output final.mp4 (-q draft for a test)
• Voice: pip install kokoro-onnx soundfile, npx hyperframes tts, npx hyperframes transcribe
• Footage: npx hyperframes init my-edit --video talk.mp4, ask for the cut list, approve, then cut
• Captions through HyperFrames, or transcribe --to srt

Sources

Checked on 2026-11-07 and 2026-11-08. Command output is from real runs on a Mac (HyperFrames CLI 0.8.226, FFmpeg 8.1.2, Ollama 0.40.14).

• Claude pricing
• Claude Code setup
• HyperFrames
• HyperFrames with Claude Code
• HyperFrames CLI reference
• FFmpeg legal (licence)
• FFmpeg download
• FFmpeg filters: silencedetect, afade, atrim, subtitles
• FFmpeg concat demuxer
• Ollama launch announcement
• Ollama + Claude Code integration
• Ollama context length

Make full explainer videos

Want an agent that turns a topic into a full explainer like this one, with the script, the voice and the scenes? That is the explainer-video agent: browse the AI agents.

Get the next free guide

New step-by-step guides and coding tutorials, straight to your inbox.