Orkas-VideoStudio pins a vendor render engine and ships no release yet
Turn your coding agent into a video studio: describe a video in plain language, and your agent writes the timeline and produces the file.
At a glance
- What is it?
- Orkas-AI/Orkas-VideoStudio lets a coding agent drive video work by writing a readable plan file rather than calling a black box. The interesting parts are its structure: three capability lines with no paid keys on two of them, a delivery guard that checks the finished cut against what was promised, a render path that is a thin wrapper over another company's package, and an npm install line the documentation itself says does not work yet.
- Who is it for?
- Use it if you want video work your agent can edit rather than re-prompt from scratch, since the plan file is the actual product here and the delivery guard is the part that gives you confidence a render matched the brief. Do not adopt it on the assumption the packages are installable: the documentation states the npm releases are still being published and tells you to build from source, and there is no release tag on the repository at all.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 11 days ago.
- What is it written in?
- Mainly TypeScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 13, 2026, and from our analysis. They are not legal advice.
Star history
- Orkas-AI/Orkas-VideoStudio349 ★ · -137
| Repository | 2026-09-11 | 2026-11-13 |
|---|---|---|
| Orkas-AI/Orkas-VideoStudio | 486 | 349 |
-137 Stars in the last 63 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
A plan file is the product, and the agent is only the planner
The central design decision is stated as a contrast. This is not a black-box video agent. A video is expressed as a readable, diffable, re-renderable plan file that both the agent and a human can edit, and the promise is that changing one line re-renders only that piece.
The division of labour follows from that. The agent is described as the brain, and the project ships three other things instead of a model: knowledge, meaning what makes a good video and which production line to take; deterministic capabilities, meaning render, edit, transcribe and generate, which are described as thin wrappers over a rendering package, the media tool, and a speech-to-text implementation; and the editable intermediate representation itself.
The example of the plan being a contract is the clearest statement of the idea. A request for existing footage in the middle, a generated opener, composed statistics sections, and one narration track over everything becomes a single editable file in which every segment, caption and narration line is independently re-renderable.
That is a real answer to the problem where a video tool is unusable because you cannot ask for a two-second change. Whether the intermediate format survives contact with a real project depends on how expressive it turns out to be, and the repository points to a plan document at its root as the place where that format is described.
A delivery guard checks the cut against the promise before it ships
One mechanism is worth singling out because it is unusual in this category. When a job spans more than one capability line, an automatic pipeline walks the plan and delegates each segment back to the line that owns it, and then a deterministic delivery guard runs a promise check on the finished cut before anything ships.
The check is given a concrete example of what it catches: a render that contains real motion rather than a silent slideshow. That is a failure mode that is easy to miss by eye, easy to ship by accident, and annoying to discover after publication. Catching it with a command rather than a reviewer's attention is the right kind of engineering.
The word deterministic is doing real work in that sentence. The capability wrappers are described as deterministic operations, and the guard is described as deterministic too, so the design distinguishes clearly between the part a model decides and the part that does not. In a session, the agent reads the router skill first to lock the production line, then the relevant stage skills, then writes the composition and the plan, runs the operations, and finally self-verifies with the guard.
What is not stated anywhere in the visible documentation is the list of promises the guard actually knows how to check. One example is given. Whether it can detect a wrong duration, a missing caption, a mismatched voice, or a narration that drifts out of sync is not described, so the guard's coverage is narrower than the idea suggests.
Two of the three capability lines need no paid key and one needs your own
The capability lines are described as orthogonal, and they differ sharply in what they cost you.
The compose line goes from a script to designed motion graphics to a finished file, covering explainers, kinetic typography, lower thirds, data visualisation, title cards and transitions. It is explicitly marked as requiring no paid keys. So is the edit line, which covers cutting, joining, trimming silence, removing filler words, mixing, burning in subtitles, dubbing and localising footage you supply, plus highlight selection over long recordings. The generate line, which produces talking-head or cinematic footage and imagery, is the one that needs your own provider keys, with three named providers and an explicit promise of no managed backend and no lock-in.
That split is a genuinely useful design. Most of what people actually want from a video tool is composition and editing, and neither needs a model account. The dependency on a paid provider is confined to the line where it is unavoidable.
Generation is also described as opt-in rather than merely optional, which is the right word: nothing about it runs until you ask for it. The three use cases in the table that involve AI imagery all sit on the generate line, and the one auto case that crosses lines is built around composing and narration rather than around generated footage.
The render path is a thin wrapper over another company's package
The most consequential dependency is named in the prerequisites rather than in a dependency list. Compose drafts use a quality gate backed by the rendering package at a pinned version, 0.7.60, described as a package dependency with the package runner as only a compatibility fallback.
That package lives under a different organisation on the forge, one belonging to a commercial video company. So the deterministic rendering path, the part of this project that is supposed to be reproducible and inspectable, is somebody else's artefact at a fixed version. The word thin wrapper is honest and also worth taking seriously: it means there is little logic here to audit, and most of the behaviour comes from a dependency you did not choose and cannot patch.
Pinning the version is the right instinct, and pinning it exactly is better than a range for a rendering engine. But it also means the project's own compatibility surface is bounded by that version, and the documentation's phrasing, with the package runner demoted to a fallback, suggests the gate is tightly coupled to it.
The second and third wrappers are the media tool and a speech-to-text implementation, both of which you install yourself on the path. Those are conventional. The rendering engine is the one to think about.
The npm install line is documented as not working yet
There are two installation paths and one of them is explicitly labelled as not ready. An early-development note says the packages are being published and that until then you should install from source, adding that the command line tool works exactly the same either way.
The from-source path is a clone, an install, a build, and then a doctor command that verifies the media tools and the runtime, with a commented alias line so the tool can be invoked by a short name from the repository directory:
git clone https://github.com/Orkas-AI/Orkas-VideoStudio.git
cd Orkas-VideoStudio
pnpm install && pnpm build
node packages/cli/dist/index.js doctor # verify ffmpeg/ffprobe/node
# optionally alias it: alias ovs="node $PWD/packages/cli/dist/index.js"The from-registry path is a single global install followed by the same doctor command, with a comment noting it provides the command itself.
So at the moment there is no package to install. The repository has no release tag at all, and the instructions for the registry path are written for a future that has not arrived. That is not unusual for an early project and the note is honest about it, but it does mean the quick path in the documentation cannot be followed today.
The version field in the manifest is the other loose end. It is a calendar-style version rather than a semantic one, so it sorts correctly by date and tells you nothing about compatibility.
Three ways to wire an agent in, and one of them puts files in your home directory
An agent can pick this up three different ways, and the choice determines what lands on your disk.
The first materialises the knowledge pack into a native skills directory so the agent discovers it by progressive disclosure. Two targets are offered, and both write outside the project: one goes to a directory under your home folder, with a repository scope option for a project-local path instead, and the other goes to a different agent's skills directory under your home folder. The second registers the MCP server, which mirrors the command line one to one, again through a package runner with a yes flag, with a note to point at the built server file directly until the packages are published. The third needs no loader at all: the command line tool can print a skill's full instructions into context, list what is available, or print its own help.
The first option is the one with a lasting footprint, since it copies instruction files into a per-user directory that applies to every project. The third has no footprint beyond the build, which is why it is presented as the option for any agent that can run a shell.
The default target for the first option is the home directory rather than the repository, and the repository scope has to be asked for explicitly. That default is the opposite of what the rest of the project does, since the whole design argues for a plan file you can diff and review inside your project.
The capability splits along payment lines, and the router hides it
Read the use-case table closely and the cost story resolves. The first case, a vertical explainer with narration and captions, routes to compose, and the narration step is marked as using a provider you supply. The second, turning an hour of screen recording into three highlight clips with captions, routes entirely to edit and needs nothing paid. The third, adding narration and subtitles to an existing file and normalising loudness, routes to edit and describes a chain that starts with transcription, goes through localisation, then speech, mixing and subtitle burn-in, ending with a loudness pass. The fourth, cleaning silences and filler words out of a webcam take, is described as a deterministic jump cut with an auditable evidence trail.
Two things stand out. The edit line is doing considerably more than its name suggests: it includes localisation and dubbing, which are speech tasks, and the only speech-to-text implementation in the stack is a wrapper, so the transcription and localisation steps ride on it. And the fourth case is the clearest statement of the project's philosophy in the whole document, because a jump cut with an evidence trail is the opposite of a model deciding what to remove.
The fifth case is a generated cinematic shot, which routes to generate and is the only row that requires a provider key. The sixth crosses lines and is guarded before delivery. So the table that is supposed to be a router demonstration is really a cost table.
Editorial conclusion
Use it if you want video work your agent can edit rather than re-prompt from scratch, since the plan file is the actual product here and the delivery guard is the part that gives you confidence a render matched the brief. Do not adopt it on the assumption the packages are installable: the documentation states the npm releases are still being published and tells you to build from source, and there is no release tag on the repository at all. Before you start, check three things: whether the pinned rendering dependency from another company is acceptable in your stack, since it is the only thing standing between you and an engine you do not control; whether your machine has the media tools the edit and transcription paths require; and whether the generation line is something you want at all, since it is the only one that needs paid keys and it is the line most likely to change your cost.
Frequently asked questions
What is Orkas-VideoStudio?
A toolkit that turns a coding agent into a video studio. You describe a video in plain language and the agent reads the material, writes a timeline as an editable plan file, and produces the file, with render, edit, transcribe and generate exposed as thin wrappers over existing media tools.
Does Orkas-VideoStudio need paid API keys?
Not for its compose and edit lines, which are both explicitly marked as requiring no paid keys. Generation is the exception: it needs your own provider keys from one of three named providers, and it is described as opt-in, so nothing on that line runs until you ask for it.
How does Orkas-VideoStudio check that a render is any good?
A deterministic delivery guard runs a promise check on the finished cut before anything ships, with a silent slideshow given as the example of what it catches. The agent self-verifies with that guard at the end of a session, and the check is described as deterministic rather than model-judged.
Can I install Orkas-VideoStudio from npm today?
Not yet. An early-development note says the packages are still being published and directs you to install from source instead, and the repository carries no release tag. The command line tool is the same either way, and a doctor command checks the media tools and runtime once installed.
What does Orkas-VideoStudio need before it will run?
Node 22 or newer, plus the media tool and its probe binary on your path, which the edit, transcribe and local quality-check paths need. Compose drafts additionally rely on a pinned rendering package, and the generation line needs your own provider keys.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/orkas-ai-orkas-videostudio)Star history chart for your README
[](https://hysenlabs.com/projects/orkas-ai-orkas-videostudio)Was this useful?