$ ./clips --auto

Automated captions and video editing

Send a video, a voice note or a line of text. Get back a finished 9:16 clip with word-by-word captions, voiceover and music, ready for TikTok, Reels or Shorts.

It runs on your own machine. No LLM calls on any run, no API bills, no per-clip cost.

Find out more See real clips

Unedited pipeline output. Tap for sound.

0: LLM calls per clip
$0: per clip: open-source models and FFmpeg
~38 s: to render an 11-16 s clip on a 4-CPU box
<10 MB: 9:16, 30 fps, loudness-levelled output

Three ways in, one clip out

Everything is sent from a phone through a Telegram bot. The finished clip comes back to the chat or lands in your TikTok drafts.

input: video

Automated edit

Speech transcribed locally with word timestamps.
Leading filler dropped, pauses over 0.8 s cut out.
Cropped to 9:16, opening hook (glitch or shake), word-highlighted captions.
Top line is your quote, or the punchiest phrase you actually said.

input: voice note

Your voice, captioned

Your own voice is kept: cleaned, trimmed and levelled.
Captions follow it word by word.
Background is a generated fractal dive, different every clip.
A new music bed is composed for each clip.

input: text

Text to video

An open neural voice (Kokoro) reads your text, on-device.
Captions are timed from the generated voice itself.
Same generated visuals and music as voice clips.
Fonts, colours, effects and music tweakable per clip.

Real clips, straight out of the pipeline

Unretouched frames from clips it rendered. No manual editing on any of them.

The caption 'JUST BELIEVE IN' in three consecutive moments of one clip, with the yellow highlight moving from JUST to BELIEVE to IN as each word is spoken — One caption line over 0.8 s: the highlight follows each spoken word.

$Text-to-video clip over a blue and gold fractal with the caption 'YOU REACH YOUR'$
$Clip over a magenta and cyan fractal with the caption 'ACQUIRE SELF-DISCIPLINE'$
$Clip over a fractal edge with the caption 'JUST TAKE ONE'$

How it works

Plain Python and FFmpeg in one Docker container. Every step is deterministic and resumable.

receiveTelegram bot takes a video (or an album, joined in order), a voice note or text. Flags like fractal=5 override the style for one clip.
transcribefaster-whisper runs locally and returns word-level timestamps.
planSegments picked from the timestamps by rules, not a model: trim filler, cut long pauses, fit the target length.
voiceYour own recording, cleaned and levelled, or Kokoro text-to-speech for typed text.
visuals9:16 crop of your footage, or a Mandelbrot dive rendered in numpy with a style and target seeded per clip.
musicA meditative MIDI bed composed per clip and rendered with FluidSynth, so no two clips share audio.
renderFFmpeg burns in word-highlighted captions and the hook, mixes to -14 LUFS, caps the file at 10 MB.
deliverBack to the chat, or uploaded to your TikTok drafts to post with one tap.

Why on-device

Free to run

Open-source models and FFmpeg. No API keys, no per-minute pricing, no usage caps. Volume costs electricity, not invoices.

Private by default

Footage never leaves your machine until you publish it. Job folders delete themselves after 30 days.

Predictable

No LLM in the loop means no prompt drift and no surprise edits. Same input, same rules, same style. A failed job resumes from the step that broke.

Want this for your content?

I can set it up for your brand, fonts and channels, or build a pipeline tailored to your workflow. Book a 30 minute call.

Find out more