AI column

Video@donghyeok99Staff

OpenMontage: your coding agent as a whole video studio

Its README shows a 60-second Pixar-style short about a lonely banana, made from six clips for $1.33 in total

A request comes in, six storyboard cards fill, a playhead crosses the timeline, the 60 s ring closes, and final.mp4 lands with a $1.33 tag.
A request comes in, six storyboard cards fill, a playhead crosses the timeline, the 60 s ring closes, and final.mp4 lands with a $1.33 tag.
Who
Built by calesthio on GitHub, nights and weekends, for people making videos from Claude Code, Cursor, Copilot, Windsurf or Codex.
When
Opened 2026-03-29, last commit 2026-11-03. No releases, you pull main. Its README wears a GitHub Trending 'Repository of the Day' badge.
Where
Inside Claude Code, Cursor, Copilot, Windsurf or Codex, on your own machine (macOS, Linux, Windows PowerShell). Cloud keys are optional.

What was hard

Most AI video tools answer one prompt with one clip. Research, script, narration, music, captions and the edit lived in other tabs, and stitching the pieces together usually produced a slideshow that moved. The bill only became clear afterwards.

What it does about it

OpenMontage gives a coding agent the entire production pipeline. There is no orchestrator code: YAML manifests and Markdown director skills teach the stages, research→proposal→script→scenes→assets→edit→compose, and the proposal, script, scene plan and assets each pause for your approval. With no keys at all it still produces a real video from Piper narration, free footage from Archive.org, NASA and Wikimedia, and a Remotion composition, then reviews its own render with ffprobe and frame samples.

The README's 60-second short "THE LAST BANANA": six clips, $1.33 in total
The Backlot board: stages light up while assets generate
The Backlot board: stages light up while assets generate
The storyboard gate: takes and cost per scene, approved before the render
The storyboard gate: takes and cost per scene, approved before the render
Script gate: the board says what it is waiting for, you answer in chat
Script gate: the board says what it is waiting for, you answer in chat
git clone https://github.com/calesthio/OpenMontage.gitcd OpenMontagemake setup"Make a 60-second animated explainer about how neural networks learn"
The README's own run

What people made with it

LinkedInKarl Weinmeister2026-06-30

Someone who hates editing tried OpenMontage: the video is built through the research→proposal→script→assets→edit→compose pipeline instead of a timeline, with a recording of the run.

So I tried out OpenMontage. It's an open-source, agentic framework that treats video production like software engineering.
likes 95 · comments 12 (2026-11-06 as of)LinkedIn ↗
GitHubdoctorrajvanshi2026-07-03

An experience report from a first full project run, attached as a Markdown file, with the fixes that got it through.

My first project experience report. Some fixes that worked.
likes 1 · comments 1 (2026-11-06 as of)GitHub ↗

How it runs inside

  1. 1

    You → Claude Code

    "Make a 60-second explainer"AGENT_GUIDE.md first: a vague ask goes through onboarding, a concrete one straight to pipeline selection.AGENT_GUIDE.md
  2. 2

    Claude Code → animated-explainer.yaml

    Read the pipeline manifestThe YAML lists the eight stages with each one's tools, review focus and success criteria, and a $2 default budget.pipeline_defs/animated-explainer.yaml
  3. 3

    animated-explainer.yaml → executive-producer.md

    Load the director skillsRuns stages in order, reviews each artifact, and sends work back, say to the script, when narration runs long.skills/pipelines/explainer/executive-producer.md
  4. 4

    executive-producer.md → tools/ registry

    registry.discover()Every BaseTool subclass under tools/ is registered; the registry says which providers work with the keys you have.tools/tool_registry.py
  5. 5

    tools/ registry → executive-producer.md

    Provider scores and cost estimateTask fit 30%, quality 20%, control 15%, reliability 15%, cost 10%, latency 5%, continuity 5%; the choice is logged.lib/scoring.py
  6. 6

    executive-producer.md → You

    Approve proposal, script, scenes, assetsNo gated stage saves as completed without recorded approval; checkpoint JSON under projects/ feeds the Backlot board.lib/checkpoint.py
  7. 7

    executive-producer.md → tools/ registry

    video_compose: Remotion renderThe runtime locked in edit_decisions picks Remotion, HyperFrames or FFmpeg; silent swaps are forbidden, so it asks you.tools/video/video_compose.py
  8. 8

    tools/ registry → final.mp4

    Scenes, captions and audio in one fileThe Explainer composition draws cards, charts, word-level captions, narration and music frame by frame into final.mp4.remotion-composer/src/Explainer.tsx

What goes in, what comes out

"Make a 45-second animated explainer about why the sky is blue"An explainer with Piper narration and Remotion animation, no keys needed (per the README)
A 60-second Pixar-style short about a lonely banana and a kiwiSix Kling v3 clips, Chirp3-HD narration, word-level captions, $1.33 in total (per the README)
"Here's a YouTube Short I love. Make me something like this, but about quantum computing."The reference's pacing and structure analysed, then 2–3 differentiated concepts with costs
A 90-second documentary montage of a city at 4am, real footage only, no narration (README prompt)Real archive footage found by CLIP search and cut into a montage, no paid model

Is it for you

Good for

  • People who post explainer videos regularly and are tired of editing
  • Anyone who wants to turn a short they love into a video on their own topic
  • Cutting a long podcast into several social clips
  • People already in Claude Code or Cursor who do not mind the terminal
  • Anyone who wants the cost per video shown first, with a cap

Not for

  • AGPL-3.0: a company may use and modify it freely, but a modified version offered as a network service must publish its source.
  • The free, keyless path covers image-based animation and archive footage. The README's examples with generated clips cost about $0.15 to $3 each in model fees.
  • You need Python 3.10+, Node.js 18+ and FFmpeg, so the install is not light. On Windows the README suggests npx --yes npm install if npm install fails.
  • Running on a local LLM (Ollama, LM Studio) is listed in the README as coming soon; today it needs a coding-agent subscription.

Sign in to post, comment and like.Sign in

Report

Reason

Read the community rules