OrkasVideoStudio

OrkasVideoStudio lets a coding agent compose, edit, and generate video from a plain-language brief. Rather than hiding the work behind a single opaque call, a video is expressed as a readable, diffable plan.json that both the agent and you can edit, and changing one line re-renders only that piece. The project supplies the production knowledge, the deterministic capabilities as thin wrappers over HyperFrames, ffmpeg, and whisper.cpp, and that intermediate representation; the agent supplies the judgement.

Four production lines cover the work. Compose turns a script into designed HTML motion graphics and then an MP4 for explainers, kinetic typography, lower thirds, data visualisation, title cards, and transitions. Edit cuts, joins, trims silence, removes fillers, mixes, burns in subtitles, dubs and localises footage you supply, and selects highlights from long recordings. Generate produces talking-head or cinematic footage and imagery through your own provider keys. Auto weaves the others together: a planning stage builds the edit decision list, an assembly stage walks it and delegates each segment, and a deterministic delivery guard checks the finished cut against its promise before anything ships.

The ovs CLI is the canonical interface and the MCP server mirrors it one to one, so an agent can pick it up three ways: installing the skill knowledge pack natively, registering the MCP server, or simply running the self-describing CLI and printing a skill's instructions into context. A session starts with a router skill that locks the production line, then reads the stage skills, authors the composition and plan, runs the deterministic operations, and self-verifies with the delivery guard.

The compose, edit, and transcribe trunk needs no paid keys. Generation is opt-in and uses your own credentials, configured by file or environment variable, with no managed backend.

Capability	Providers
Image	OpenAI-compatible, Gemini, Doubao Seedream
Video	Doubao Seedance image-to-video
Speech	OpenAI-compatible text-to-speech endpoints

Features

• Editable plan: a plan.json intermediate representation you can read, diff, and re-render per segment
• Compose line: script to HTML motion graphics to MP4, with no paid keys required
• Edit line: cut, join, trim silence, remove fillers, mix, burn in subtitles, dub, localise, and rank takes
• Generate line: image and video generation through your own provider keys
• Auto pipeline: one cross-modal plan routed across compose, generate, and narration for multi-axis deliverables
• Delivery guard: a deterministic promise check verifies the finished cut before delivery
• Agent entry points: native skills, an MCP server mirroring the CLI, or the self-describing CLI itself
• Narration fit: speech capability resolution and per-line fitting keep narration inside its plan window
• Local-first stack: rendering, editing, and transcription run through HyperFrames, ffmpeg, and whisper.cpp on your machine
• Deterministic tests: a fake provider endpoint and fixture-based checks exercise the video path without spending credits