SkillFoxx

OpenVideo skill

Local video with MiniMax H3 on your own GPU: the open-video CLI and a skill that lets an agent write prompts and run clips.

  • 117 ★ on the repository
  • Last commit
  • Apache-2.0
Skill

Extended access

Why this level

  • The installer downloads about 54 GB of weights and sets up ComfyUI, and the agent runs scripts and heavy GPU generation.
  • The VLM judge sends frames to the address in `OPEN_VIDEO_VLM_URL`, so choose a trusted service.
All reasons and checks

agent-next/video-agentAuthor:agent-next

Install

In your terminal, with SkillFoxx CLI

npx skillfoxx add skills/openvideo

Detects the agents on your machine, shows the access level and pins the version.

Other ways to install

Access and checks

How we review

Extended accessThe tool runs code or makes network calls and only reads external data.

  • The installer downloads about 54 GB of weights and sets up ComfyUI, and the agent runs scripts and heavy GPU generation.
  • The VLM judge sends frames to the address in `OPEN_VIDEO_VLM_URL`, so choose a trusted service.

This is third-party code. Review the repository files before installing.

What it does

OpenVideo is a CLI and skills for local video generation with the MiniMax H3 model through ComfyUI. The `open-video pull`, `status` and `run` commands are modelled on Ollama: fetch weights, check the engine, generate a clip. The `h3-video` skill takes an agent through the full loop: it writes a three-field prompt following the H3 grammar, validates constraints, runs the T2V, I2V or FL2VA modes and reviews the output. The `open-video` skill adds an experimental director for multi-shot clips: a plan, scoring by a VLM judge, retries and stitching through ffmpeg.

Who it is for. NVIDIA GPU owners who want clips from text, an image or a pair of frames without cloud services.

Good fit when

  • A short clip from text: the agent writes the prompt and runs `open-video run` with the duration.
  • A clip from an existing image or from first and last frames.
  • Planning and parameter validation without spending GPU time: `open-video "..." --dry-run`.
  • `open-video recommend-quant` suggests a quantisation tier for a specific card.

Not a fit when

  • There is no NVIDIA GPU and neither Linux nor WSL2 is available: the main H3 generation path runs through them.
  • A clip longer than 15 seconds without manual work: multi-shot planning is experimental.
  • The H3 weights licence does not fit your region or commercial use.

Example request

Make an eight-second clip of an astronaut planting a flag on a red dune at dusk. Validate the prompt with a dry run first.

Limitations

Installation downloads about 54 GB of weights. The main INT8 tier targets cards with 22 GiB of video memory or more, ComfyUI offload kicks in at 12–22 GiB, and the W4 and NF4 tiers are experimental and installed by hand. The code is Apache-2.0, while the H3 weights fall under the MiniMax H3 Community License with territorial and commercial limits. The VLM judge turns on with `OPEN_VIDEO_VLM_URL` and `OPEN_VIDEO_VLM_MODEL`, and without them scoring is skipped. Version 0.1.0 generates single clips.

How to disable. Remove the skill from the agent's skills folder and delete the cloned project folder. Weights and the ComfyUI installed by the script are removed separately.

FAQ

How do I install OpenVideo?

Clone tag `v0.1.0` and run `bash scripts/install.sh`, which sets up ComfyUI and the H3 weights. Then `open-video pull h3` and `open-video run "..."`.

Which skill do I use for ordinary clips?

`skill/h3-video`, the main one in version 0.1.0. `skill/open-video` with planning, judging and stitching is marked experimental.

Does generation work on macOS?

Installation and dry runs work. H3 generation on macOS goes through community and MLX paths and is not the default route.

How do I enable judge scoring?

Set `OPEN_VIDEO_VLM_URL` and `OPEN_VIDEO_VLM_MODEL` for an OpenAI-compatible vision model. Without them scoring is skipped and reported as such.

How much video memory is needed?

The main INT8 tier targets 22 GiB and up, and ComfyUI offload engages at 12–22 GiB. `open-video recommend-quant` picks a tier for your card.

Discussion

No questions or reviews yet.

Skills

The skill retells a reference video, waits for your confirmation and writes MiniMax H3 prompts per scene and shot.

SkillExtended access325
Official

The official mmx CLI of the MiniMax platform with an agent skill: text, images, video, speech, audio transcription, vision and search

CLIExtended access2.2K

Codex skill that makes short MiniMax H3 videos from your images on a Google Colab GPU, with a job queue and result download

SkillExtended access136

A skill and CLI that let Claude Code or Codex make videos as web pages, with the Studio editor to change the result

SkillExtended access5.9K

A CLI for programmatic video from HTML: the agent assembles frames from templates and renders a finished MP4 on your machine

CLIExtended access4.8K

A skill that pulls highlights from a long video: download, transcript, segmentation, analysis and clips with titles

SkillExtended access579

Comparison with similar tools

Cards that share the main search query, and how they differ.

ToolOfficialRepository starsMain difference
OpenVideo (this card)no117
HyperFramesno83,283renders MP4 from HTML compositions; this one runs local neural generation with H3
video-useno30,939edits existing recordings by cutting pauses and adding subtitles; this one creates the clip from scratch with a model

README in short

The chain runs from prompt or skill to the `open-video` CLI, the H3 backend, the ComfyUI engine and a finished mp4. Multi-model backends and a full long-film mode are listed in the README as future work.

README badge

Original SKILL.md

---
name: h3-video
description: >
  OpenVideo skill (v0.1.0): generate high-quality local video with the OpenVideo product
  (MiniMax H3 backend). Use for OpenVideo install/pull/status/run, official 3-field prompts,
  T2V/I2V/FL2VA, agent-driven video. Brand is OpenVideo — not a bare ComfyUI workflow.
  Triggers: OpenVideo, open-video, H3, generate video, T2V, I2V, FL2VA.
---

# OpenVideo skill · v0.1.0

**Brand: OpenVideo** (always lead with this). MiniMax H3 is the model OpenVideo drives.

**Job:** use **OpenVideo** so an agent produces **good product video**, not a random diffusion
click. Quality comes from **prompt craft + correct mode + validated settings**, not secret samplers.

## 0. Paths (env-first)

Recommended sibling layout (any parent dir name; no machine-absolute paths):

```text
parent/
├── open-video/       # THIS product — CLI, backends, skills  ← work here
└── lab/              # ComfyUI + weights (not git)
```

| What | Path / env |
|---|---|
| **Product root (CLI, skills)** | `$OPEN_VIDEO_ROOT` (this checkout) |
| **Install / pull / run** | From product root: `./scripts/install.sh`, `python -m open_video …` |
| **Lab engine + weights** | `$OPEN_VIDEO_LAB` / `$H3_LAB` (sibling `lab/` with ComfyUI + `h3_models/`) |
| **Weights env** | `export OPEN_VIDEO_MODELS=$H3_LAB/h3_models` |
| **ComfyUI URL** | `export OPEN_VIDEO_COMFYUI=http://127.0.0.1:8188` (default) |

**Resolve product root** (never hardcode a machine path):

```bash
# Prefer env, else directory of this skill → repo root
export OPEN_VIDEO_ROOT="${OPEN_VIDEO_ROOT:-$(cd "$(dirname "$0")/../.." 2>/dev/null && pwd)}"
cd "$OPEN_VIDEO_ROOT"
```

Lab harness (same ComfyUI) when you keep a sibling runtime:

```bash
export OPEN_VIDEO_LAB="${OPEN_VIDEO_LAB:-$OPEN_VIDEO_ROOT/../lab}"
export H3_LAB="${H3_LAB:-$OPEN_VIDEO_LAB}"
export OPEN_VIDEO_MODELS="${OPEN_VIDEO_MODELS:-$H3_LAB/h3_models}"
```

Report a problem