Everyframe, a video editor built on Renderbox

01 · The app

You describe the video. Everyframe makes it.

Everyframe is an online video editor with an AI assistant. It works from your own clips. The assistant shows you the plan and the cost first. Nothing is made until you say yes.

Everyframe

AI assistant

Describe a task. Everyframe plans the work, shows the cost and helps you get it done.

Composer

Video ads

Choose your clips, words and voice. Build video ads and make variations in one place.

Cut

Timeline editor

Trim, arrange and layer your clips. Add text, sound and music. Use AI tools to speed up repetitive edits.

Library

Your media

Keep your clips, images and voices together. Use the same media across your editing tools.

Talk to Everyframe ↗ everyframe.si

02 · How it uses Renderbox

Every render runs on Renderbox.

Everyframe does not render video itself. It sends each export, caption and clip analysis to Renderbox as a job. It builds the jobs with @jtdigital/renderbox-sdk.

Everyframe keeps your projects and files. Renderbox runs the encoding, the speech recognition and the AI models.

Each job reports its progress, then ends as completed or failed. If a result arrives twice, Everyframe saves it once.

Every minute, Everyframe marks the jobs that never answered as failed.

# out, one job per output file
job.submit                       graph_ir + bindings, job_id from the row

# back, on queues named for its tenant
results.<tenant>.progress
results.<tenant>.completed       settled by job_id, once
results.<tenant>.failed          transient · permanent · fatal · compile
results.<tenant>.manifest
results.<tenant>.dead            what it cannot read

Names from everyframe-composer's src/infra/renderbox. It copies the engine's message types into TypeScript.

03 · Computer vision

Renderbox reads each clip once.

When you press Analyse on a clip, Renderbox finds the faces, the shots, the words, the music and the sound. Everyframe saves the results with the clip. Every tool reuses them, so no clip is analysed twice.

FACES: detect, track; yolov8n-face
SHOTS: scene labels; places365-resnet50
WORDS: transcribe; parakeet-tdt-0.6b-v3-int8
MUSIC: visual_peaks, analyze_audio; no model, the engine's own measures
ROOM: pan, aresample, write_audio; no model, FFmpeg only

It follows the face that matters.

Uses the FACES read.

It cuts on the words.

Uses the WORDS read.

Every cut lands on a beat.

Uses the MUSIC read.

One brief, every version.

27 files, one job each.

04 · Animation

Animations are written as code.

Everyframe's assistant writes each animation in TypeScript with the Renderbox motion API. Renderbox renders it.

Text and numbers come out exactly as you typed them. The same animation renders in every size and language. No AI image or video model is used, so there is no generation cost.

Everyframe can make 11 of its 16 example animations today. The other 5 need the new renderer. It is not on the servers yet.

The motion API and renderer →

launch-board · the bars grow and swap places, and the total counts up

05 · The assistant

Every video is checked before you get it.

The assistant works in a throwaway container with your clips. The container holds one key, for the AI model. The keys for storage and paid services stay outside it.

Four checks run before it hands a video back. If a check fails, the assistant fixes the problem and runs the checks again, up to two times. If a check still fails, the job ends as failed.

The checks, on the Motion page →

01 rbv check the motion code compiles
02 probes each clip has the right shape, frame rate and loudness
03 critic a separate AI reviews the frames, and the assistant cannot change its verdict
04 hand-back the final file passes the app's own rules

Everyframe is at everyframe.si.