Automate AI filmmaking with the Venice Video Harness

What the Video Harness is and what it produces

The Venice Video Harness is open-source software that lets you and an AI agent build a short film from a single idea, using nothing but a Venice API key. Instead of manually prompting shot after shot in Venice Studio, you hand the harness your premise, genre, setting, and goals, and it runs the agent through the entire film: story beats, reference images, video generation, and a final assembly you can watch.

Set expectations honestly. The first pass will not be perfect, but it gets close and does the bulk of the busy work. What you get is a solid starting point to polish in the editing room, not a finished master. For example, a comedic detective story set in Renaissance Venice was produced in a few hours this way.

Installing and running from the command line

The command line is the most advanced route, so skip it if that sounds intimidating. Open a terminal and install the software, then run the setup command to register your API key.

npm install flagg-venice-video-harness
venice-video-setup

Setup asks for your Venice API key. Grab it from the Venice API settings page and paste it in. It stays hidden on screen for security. Choose a default project workspace, and you are ready.

To start a film, run the new-film command and answer the walkthrough prompts:

venice-video-new

You will name the film, write a one-line premise, and set genre, setting, and era. Then you pick an audio strategy: native dialogue generated by the video model, lip sync from separate audio files, or narrative voiceover. You choose a video model family (the default is Seed Dance), and because the CLI is not running an agent, you also pick an intelligence model to reason through the story. From there the harness walks you through the whole narrative, generates the reference files and videos, and assembles everything into a watchable cut plus an editable XML file.

Driving the harness with the OpenCode agent

If the command line is too much, use a coding agent instead. Download OpenCode from opencode.ai as a desktop app, then open its settings, find the Providers tab, and choose Venice. Add your API key to the new Venice provider. The Models tab will show every model available on Venice through the API, and you can toggle individual models on or off.

Model choice affects both quality and cost. Fable tends to give the best results but is expensive and not private. Cheaper strong options include Kimi K3 fast and GLM 5.2, and Sonnet or Opus also work well. Check current pricing at docs.venice.ai before you commit. For example, Kimi K3 was priced at 450 per million input tokens and 2250 per million output tokens, while Fable was $12 per million input and $60 per million output.

Once your model is selected, create a new project (a new folder on your computer), open a new session, copy the harness GitHub repo URL, and tell the agent to set it up:

Let's set up the Venice Video Harness.

A note on privacy: if you use a non-private model like Fable, revoke that API key on the Venice API page when the project is done by clicking the trash icon. That prevents a leaked key from being reused.

Running a film from a single brain dump

With the harness installed, you do not need structured commands. Just describe what you want in plain language and let the agent take over. A basic brain dump is enough:

Let's use the Venice Video Harness to create a short film about Venice
hitting four million users. A detective in Renaissance-era Venice is trying
to find dirt on the city's citizens but can't find anything, because Venice
keeps no logs. Make it comedic, fun music, intriguing and curiosity-provoking.
Walk me through it, then create the video.

Because the harness is installed, the agent knows where to look and which rules to follow. It thinks through the concept, story beats, characters, production plan, and dialogue strategy before asking you a short list of questions: frame format (for example 16:9 cinematic), how the story should be voiced, how narration should sound, and whether to burn subtitles into the final video. Answer those, and when it runs out of questions it generates the film.

Fixing shots and exporting an editable timeline

Expect a few shots to miss on the first pass, often a wrong reference image or an inconsistent object. There are three ways to fix them.

The fastest is to ask the agent in plain language. For example, tell OpenCode that a shot is inconsistent, then instruct it to regenerate the reference image and re-render that shot:

In shot 5 the machine is not consistent. Regenerate the reference image of
the note coming out of the machine, then regenerate that shot.

The second way is manual, through Seed Dance. Find the folder the harness created, go into episodes, open your episode, and open scene one to see every video file. Upload the problem shot as a reference to Seed Dance 2.5, use the edit-video prompt, and pull correct references from the character folder in the media library. Specify the exact change, the shot length (for example five seconds), keep the aspect ratio, and submit.

The third way is for advanced creators. The harness also outputs an FCP XML file inside its working folder. Import that XML into DaVinci Resolve, Premiere Pro, or Final Cut Pro and every shot lands on your timeline with audio attached. From there you can detach audio for L or J cuts, adjust levels, tighten pacing, and do granular edits in the editor you already know. It is still a work in progress, so you may see small gaps to clean up.

Previewing with the browser workshop

The harness includes a visual browser workshop, still in development, that lets you inspect what the agent is doing before it commits to generating video. Ask the agent to launch it so you can review reference images ahead of rendering:

Run the browser workshop so I can see the reference images before you
generate any videos.

The agent opens it in your browser and you can look at everything the harness is preparing. This is a community project, so feedback is welcome in the Venice Discord. The larger point holds: all you need is a Venice API key, and when you choose a private model there is zero data retention, so nothing you send gets logged anywhere.

Key takeaways

• The Venice Video Harness turns a single premise and one API key into a rough-cut film, handling story, references, shots, and assembly.
• You can run it three ways: the command line, the OpenCode agent, or a plain-language brain dump once it is installed.
• Model choice trades cost, quality, and privacy; check docs.venice.ai for pricing and revoke non-private keys when you finish.
• Plan to polish the output by re-prompting the agent, editing shots in Seed Dance, or importing the XML timeline into your own editor.