OpenVideo v0.1.0 · Apache 2.0

OpenVideo
open video, on your GPU

OpenVideo is open-source video generation: one-line install, simple pull / run for MiniMax H3, and a skill harness so any agent can produce high-quality video on your GPU. Works today; GitHub when we open it.

Install free → Try the director UI Docs

terminal · open-video install NVIDIA GPU

$ curl -fsSL https://open-video.ai/install | bash

OpenVideo local-first MiniMax H3 agent skill

demo.mp4 · director cut

00 Same pattern across creative work

OpenCode open · code agent ↔ Cursor Open Design open · design workspace ↔ Claude Design

OpenVideo OpenVideo ↔ Runway / Seedance

We take install UX + open-core from OpenCode, local agent product from Open Design, and creator-facing product DNA from OpenArt — applied only to video.

Apache 2.0 open core · Local-first on your GPU · Model: MiniMax H3 · Arena #1 open · Engine: ComfyUI · Agent hosts: Claude · Cursor · OpenCode · skills

01 The thesis

Open video models already match closed quality on short clips. OpenVideo v0.1.0 is local MiniMax H3 + an agent skill harness — install, pull, run on your GPU. Apache 2.0, local-first. The VLM judge with refine retries is opt-in; the multi-shot film director is experimental (scaffold, not finished product).

01 install → 02 pull H3 → 03 run roadmap: judge → refine → stitch

02 Interfaces

What works today. What’s planned.

v0.1.0 leads with CLI + agent skill on local H3. App, HTTP API, and MCP are roadmap — not current product surfaces. /try is a browser mockup only.

CLI

Ships today

Install, pull H3, status, run. Pipe prompts on your NVIDIA GPU via ComfyUI. Dry-run validates without a full render. No hosted free GPU.

Skill

Ships today

Packaged skill for agent hosts (Claude Code, Codex, Cursor, OpenCode, …). Official H3 3-field prompts — high-quality local clips, not one-liners.

App / API / MCP

Planned

Desktop app, HTTP API, and MCP are design targets. Not available as product surfaces in v0.1.0. Use CLI + skill (and the /try mockup for UX feel).

03 Built on open models

Standing on the shoulders of open giants.

No closed dependencies. Every layer of the stack is open weights and open source — auditable, hackable, and yours to run locally.

ComfyUI

node runtime

MiniMax H3

video model

VideoScore

judge (roadmap)

woodfantasy

open weights refpack

04 Quickstart

Install like OpenCode. Film on your GPU.

Same one-liner pattern as opencode.ai/install and Open Design. No sign-up. Local NVIDIA GPU required for real renders.

bash — open-video

1 $ curl -fsSL https://open-video.ai/install | bash

2 $ open-video "a neon koi swimming through rain, slow dolly"

? # needs NVIDIA GPU · ~60GB disk for first download · see docs docs

05 Open vs closed

Why not just use Runway?

Same job as the closed products — concept to film — without the meter, the watermark, or the black box. Pattern learned from OpenCode vs Cursor.

	OpenVideo	Runway / Seedance
Runs on your GPU	Yes · local-first	No · cloud only
Cost for unlimited local use	Free (Apache 2.0)	Per-second / subscription
Director loop (judge → refine)	Design / scaffold	Closed · internal
Model choice	Pluggable (H3 today)	Vendor-fixed
Region / watermark locks	None on your machine	Common
Agent skill / CLI	First-class today	Limited / proprietary