kineto

An agent-first video studio built on Remotion.

Describe a video in plain words. Claude Code or Codex builds it through one CLI,
and every asset and render stays on record.

kineto promo preview: Clawd flies through rings on a rainbow trail, then a quill writes kineto

_{▶ The preview loops without sound. Watch the full 55-second promo, made in this repository with kineto.}

English · 简体中文

✨ Why kineto

AI-made videos are everywhere, yet the way they are made is rarely visible.
Coding agents already write Remotion code. Left alone, they guess outdated APIs, scatter files and lose track of what they rendered.
kineto gives the agent one CLI as the only way to change records. The creative work, scenes and timelines, stays in plain Remotion.
You ask in natural language and get a real MP4. This is vibe coding applied to video.

🚪 One door

Every record changes through ./kineto
JSON output, stable exit codes, errors with a next step
./kineto check catches hand edits in CI

🏷️ Provenance for every asset

Content-addressed files with license, author and source; files without a license are refused
Generated voiceovers and captions also record how they were made

🧾 Every render on file

Videos, scenes, stills and GIFs
Git sha, codec and Remotion version per export
A shareable link when storage sits behind a CDN

🔌 Pluggable

Storage: local, or any S3-compatible service
Tools: voiceover and captions, each enabled in kineto.config.yaml

🚀 Quick start

git clone https://github.com/zephyr4123/kineto.git && cd kineto
npm install
./kineto setup        # install the official Remotion agent skills
./kineto doctor       # check Node, dependencies, config, storage and skills
./kineto new hello --title "Hello kineto"
./kineto studio hello # preview while the scenes are being written
./kineto render hello

With an agent, open it in the repository and ask for a video. It reads AGENTS.md and works from there.

🧭 How it works

person        "make a 30-second launch clip with a voiceover"
  ↓
agent         reads AGENTS.md and the official Remotion skills, issues CLI calls
  ↓
kineto CLI    the only door: JSON output, stable exit codes, errors with hints
  ↓
kernel        video catalog · asset library · render · sync · check
  ↓
strategies    engine: Remotion · storage: local, S3-compatible · tools: tts, transcribe

	What it holds	Where
🎬 Video	Code, metadata and render history	`videos/<id>/`
🖼️ Asset	Content-addressed file with license, author and source	`assets/manifest.jsonl` + storage
📦 Render	Each export with git sha, codec and Remotion version	`videos/<id>/renders.jsonl` + storage

Each render bundles only its own video, so a broken video cannot block the others.

☁️ Storage

Assets and renders stay out of git. They go to the backend set in kineto.config.yaml (start from kineto.config.example.yaml).

local is the default: a directory inside the repository, no setup required.
s3 works with Tencent COS, AWS S3, Cloudflare R2, Alibaba OSS and MinIO. Only the endpoint changes.
Objects are immutable and uploaded with their content type and a long-lived cache header.
With publicUrl set to a CDN domain, render records carry a shareable link. The promo above is served this way.
Secrets appear only as ${ENV} references. envFile can point to a chmod 600 file outside the repository.
After switching to a shared backend, ./kineto storage push uploads what this machine already has.

🧰 Tools

Tools are plugins in tools/ that produce media outside Remotion. Each one is named after its capability, and none runs until it is enabled in kineto.config.yaml.

Tool	Input → output	Provider
🗣️ `tts`	Text → voiceover WAV	Tencent Cloud TTS
💬 `transcribe`	Speech in an asset → captions for `@remotion/captions`	whisper.cpp, local and free

./kineto tool tts "Every video starts with a single line." --to hello --as voice
./kineto tool transcribe voice --to hello --as captions

Results enter the asset library with license and provenance, like any other asset.
Settings and --to / --as are checked before a paid API is called.
A new plugin is one file. See tools/README.md.

⌨️ CLI

./kineto help

Command	Purpose
`doctor`	Check that this machine can make videos
`setup`	Install the official Remotion agent skills and sync generated files
`new <id> --title <text>`	Create a video from a template and register it
`list` / `show <id>` / `update <id>`	Browse and edit the catalog
`asset add <file\|url> --license <spdx>`	Register an asset, optionally with `--to <video> --as <alias>`
`asset link <asset-id> --to <video> --as <alias>`	Reuse a registered asset in another video
`asset unlink <alias> --from <video>`	Remove an asset from a video once its code no longer uses it
`asset update <asset-id> --license <spdx>`	Correct a registered license or author
`asset list`	List all assets, or those of one video
`tool list` / `tool <name>`	List tool plugins, or run one and register its result
`studio [id]`	Open Remotion Studio with every video, or with one
`render <id>`	Render, store and record a video, a scene, a GIF or a `<Still>`
`sync` / `check [--deep]`	Regenerate derived files / validate the whole library
`storage push`	Upload this machine's assets and renders to the shared backend

Output is JSON whenever stdout is not a terminal.
Exit codes: 0 success, 1 error, 2 usage error.
Commands that change records take a repository lock, so parallel agents are safe.

🛠️ Development

npm run verify    # typecheck, lint, tests and ./kineto check --deep (the CI gate)
npm run ship      # land the current branch on main once CI is green

The CLI and kernel are TypeScript run directly by Node's type stripping. There is no build step.
main only takes commits that passed CI on a branch, fast-forwarded by npm run ship. Branch, commit and changelog rules are in AGENTS.md.
Releases are published by pushing a version tag such as v1.0.0. See CHANGELOG.md.

📄 License

kineto's own code is released under the MIT License.
kineto is built with Remotion, which has its own license: free for individuals and companies of up to three people, a company license otherwise. See remotion.pro/license. These terms apply to anyone using kineto.
Each asset's license is recorded in assets/manifest.jsonl.