OpenMontage lets your coding agent direct real video, not just animate stills

OpenMontage is an agentic video production system that treats your AI coding assistant — Claude Code, Cursor, Copilot, Windsurf, Codex — as the actual orchestrator. There is no hidden code scheduler running the show. The agent reads YAML pipeline manifests to know what stages exist, reads Markdown skill files to know how to execute each one, calls Python tools for the actual work, and checkpoints state as JSON so runs are resumable. That architecture is genuinely different from wrapping a single generation API in a chat loop.

The distinction the README draws most sharply is between image-based video and real video. Most "free AI video" stacks animate a handful of stills and call it done. OpenMontage can do that, but it can also build a finished edit from actual motion footage pulled from Archive.org, NASA, Wikimedia Commons, Pexels, and Pixabay — semantically ranked via CLIP, cut into a timeline, and rendered through FFmpeg. The documentary montage pipeline is the clearest expression of this: prompt for a tone poem about city life at 4am, specify real footage only, and the agent builds a corpus, retrieves clips, and edits them rather than generating anything.

Quality enforcement is unusually explicit. A pre-compose validation gate blocks the render if the delivery promise is violated — say, a "motion-led" brief where the plan is mostly stills. Post-render self-review runs ffprobe validation, extracts frames at four positions to check for black frames and broken overlays, and analyzes audio levels. If that review fails, the video is not presented to you. A slideshow risk score across six dimensions prevents animated-PowerPoint outputs. Human approval gates are enforced at proposal, script, scene plan, and asset review — the checkpoint writer rejects a stage marked complete without recorded approval.

Provider selection runs through a seven-dimension scoring engine: task fit carries the most weight, followed by output quality, then control features and reliability, with cost efficiency, latency, and continuity rounding out the score. The winning provider and all alternatives are logged in a decision trail you can inspect. Budget governance is built in at the same level — cost estimation before execution, reservations before API calls, hard caps, and per-action approval thresholds.

The zero-key path is real. Piper TTS handles narration offline. Remotion (React-based) and HyperFrames (HTML/GSAP) handle composition. FFmpeg handles everything else. The animated short "The Last Banana" — a Pixar-style 60-second piece with generated motion clips, narration, music, and word-level captions — cost $1.33 total with paid providers configured. The system picks between Remotion and HyperFrames at proposal time and locks that choice; swapping runtimes silently mid-production is treated as a governance violation.

The scope is ambitious: over 100 registered tools, more than 60 provider integrations, and hundreds of agent skill files covering pipeline direction, creative technique, and provider-specific knowledge. Whether that breadth translates to reliable outputs in practice is something the README cannot answer — but the governance layer at least makes failures visible rather than silent.