calesthio/OpenMontage

OpenMontage is your AI coding assistant acting as a video production crew. You describe what you want in plain language; the agent handles research, scripting, asset generation, editing, and final composition. The key architectural claim is that there is no separate code orchestrator — your existing AI coding assistant (Claude Code, Cursor, Copilot, Windsurf, Codex) reads YAML pipeline manifests and Markdown skill files, then calls Python tools directly. The agent IS the orchestrator.

The distinction the README most wants to make is about real footage versus animated stills. Most "AI video" tools animate a handful of images and call it done. OpenMontage can do that, but it can also build a CLIP-indexed corpus from Archive.org, NASA, Wikimedia Commons, and free stock sources, retrieve actual motion clips, edit them into a timeline, and render a finished piece. That matters if you want documentary texture without paying for video generation APIs.

When you do want generated video, the provider roster is wide: Kling, Runway Gen-4, Google Veo 3.1, MiniMax, Seedance, Higgsfield, and local GPU options like WAN 2.1 and Hunyuan. Every provider selection runs through a 7-dimension scoring engine — task fit weighted at 30%, output quality at 20%, then control features, reliability, cost efficiency, latency, and continuity filling out the rest. The winning choice and all alternatives are logged in a decision trail you can inspect.

Quality enforcement is more serious than most open-source video tools bother with. Human approval gates pause execution at proposal, script, scene plan, and asset review — and the checkpoint writer actively rejects a stage marked complete without recorded approval. A pre-compose validation step blocks the render if the delivery promise is violated (say, a "motion-led" brief that would produce 80% stills). After render, ffprobe validation, frame sampling at four positions, and audio level analysis run automatically; the video is not presented if that review fails.

The Backlot board is a local dashboard that fills itself in as a production runs — stages light up, scene cards shimmer during asset generation, and every provider decision and dollar spent appears on the wall. Asset generation pauses on a scene-by-scene contact sheet so you approve visuals before the render. A replay feature lets you scrub the entire production run from its timestamps after the fact.

Budget governance is built in: cost estimates before execution, per-action approval above a configurable threshold (default $0.50), and a total cap defaulting to $10. The demo videos in the README show what this costs in practice — a 60-second animated short ran $1.33, a 50-second transformation film ran about $4, and a seven-world showcase ran about $5.

The zero-key path is genuinely functional: Piper TTS for narration, Archive.org and Wikimedia for footage, Remotion for React-based composition, HyperFrames for HTML/GSAP motion graphics, and FFmpeg for everything else. You can render a real edited video without a single paid API call.