OpenMontage vs Orkas:
assemble the pipeline yourself, or install one.

Yes. Orkas has a generation line that reaches Seedance 2.0, Kling 3.0 Turbo, Veo 3.1, Runway Gen-4.5, Hailuo 2.3 and Vidu Q3 Pro, and it can run them on an Orkas-managed tier or on your own provider key. The difference is not which models exist but how you start: OpenMontage expects you to register with each provider and configure the keys yourself, while Orkas can produce a generated shot on the managed tier straight after sign-in.

Both are open source and both run on your machine, and they overlap on the generation models. The difference is shape and licence: OpenMontage is an AGPLv3 repository you assemble and drive from a coding assistant, Orkas is an MIT desktop application you install. The honest limits are that Orkas cannot install into a coding assistant you already run, and its pipeline is not opened file by file.

An ordinary rendered video is not covered by AGPLv3 merely because OpenMontage produced it. The licence says output is covered only if its content constitutes a covered work. Where the licence does matter is if you modify OpenMontage and then redistribute it or offer it to others over a network; the AGPL obligations apply to that software. Orkas is MIT, which carries no such condition. This describes licence scope, not ownership, and is not legal advice.

Neither charges for the software itself; the cost is the model usage behind it. A published third-party review of OpenMontage puts a 60-second video at 200,000 to 500,000 tokens of agent reasoning — roughly 3 to 7 US dollars at Claude Opus pricing, before asset costs — while the project's own materials quote a much lower figure. On both sides, generated shots add the video model's own per-clip charge on top of that, and composed HTML scenes do not. Your own number depends on how much of the video is generated footage and which model you point either one at.

Orkas has a dedicated method for it. One front portrait is generated once and then locked as that character's anchor and never regenerated; every later shot picks its references by rule, preferring a recent frame from the same camera angle, then an older frame, then the portrait alone. Motion is described by visible features rather than by name, because the model conditions on pixels and not on labels. You can also upload a photo to become that anchor directly. OpenMontage also documents character-consistent series built from reference images and provider-specific identity tools such as Soul ID. The difference is orchestration: Orkas fixes one portrait as the character anchor and applies an explicit cross-shot reference order; OpenMontage gives the coding agent provider guidance, so the exact continuity method varies with the selected provider.

Both address it, in different shapes. OpenMontage governs a budget: it estimates before execution, reserves funds before the call, pauses above a per-action threshold, caps the total, and offers observe, warn and cap modes. Orkas gates each attempt instead: before any billable generation it states the exact shot and character count together with a credit quote and waits for approval, defaults to six shots and three characters per video, skips generating an end frame a shot does not need, and does not charge for a call that was never sent to the provider. Neither is strictly better — one guards a pool, the other guards each transaction.