Most AI video tools ask you to log into a dashboard. OpenMontage asks you to git clone a repository and let your coding agent take the director's chair.
A Repo That Plays Director, Not a SaaS Dashboard
OpenMontage is an open-source, agentic video-production framework that ships no orchestrator of its own — the AI coding agent you already run becomes the director. At the live 2026 snapshot, calesthio/OpenMontage showed 45,042 stars, 5,540 forks, and 323 commits, licensed AGPL-3.0. The repo was created on 2026-03-29 and carries no tagged releases — a rolling main branch, not a versioned product, which makes exact version claims unreliable.
The architecture is the point. Python supplies the tools and JSON/YAML state persistence; the coding agent — Claude Code, Cursor, GitHub Copilot, Codex, Windsurf, or OpenCode — owns the entire control loop. Production is encoded as a seven-stage sequence: research → proposal → script → scene_plan → assets → edit → compose. Each stage pairs a YAML manifest in pipeline_defs/ with a Markdown director skill in skills/pipelines/, as documented in the pipeline walkthrough.
"Every video request must go through a named pipeline; ad hoc scripts and bypassing the checkpoint loop are disallowed by design." — OpenMontage AGENT_GUIDE (source: AGENT_GUIDE.md)
That constraint is enforced, not merely advised: the agent reads the manifest, runs preflight, discovers tools through a registry, self-reviews, checkpoints state, and halts for human approval at creative decision points.
Before You Begin: Python 3.10, FFmpeg, and an Agent
OpenMontage runs locally by default, so the prerequisite list is short and the setup is one command. You need Python 3.10+ for tool execution and JSON state persistence, FFmpeg for post-production compositing, Node.js 18+ for the Remotion renderer, and one of six supported coding agents — Claude Code, Cursor, GitHub Copilot, Codex, Windsurf, or OpenCode — to act as director .
After git clone and cd OpenMontage, a single make setup installs all three dependency layers at once: Python packages, Remotion's Node packages, and Piper TTS for offline narration . There are no manual pip or npm steps to chase.
Copying .env.example to .env is optional — do it only to wire in paid providers such as ElevenLabs, Kling, or fal.ai. Leave it absent and you stay on the zero-key path. No Docker, no cloud provisioning, no dashboard login.
How Claude Code Directs a Video, From Empty Folder to Draft
With setup done, two terminal commands are all you type before handing off. Clone the repository and enter it:
git clone https://github.com/calesthio/OpenMontage.git
cd OpenMontage && make setup
From here, the coding agent takes the director's chair. Open the repo in Claude Code and it reads CLAUDE.md, which points to AGENT_GUIDE.md for shared operating rules . Cursor, Copilot, Codex, and Windsurf each read their own platform markdown file that routes to the same guidance . The agent then discovers what tools it can call through the registry:
python -c "from tools.tool_registry import registry; registry.discover()"
The AGENT_GUIDE is prescriptive: the agent reads the relevant skill files in skills/ before invoking anything, and every request must run through a pipeline — ad hoc scripts and bypassing review are explicitly disallowed .
Now issue a plain-English brief, such as "Make a 60-second animated explainer about how neural networks learn." The agent selects a pipeline from the staged workflow — research → proposal → script → scene_plan → assets → edit → compose — runs preflight checks, self-reviews, checkpoints JSON state, and halts for your approval at every creative decision point before proceeding .
Because the whole production is reproducible from repo state, nothing hides in a black-box render. The script lands as Markdown, scene metadata as JSON and YAML, assets under stable names, plus render commands and self-review notes — all ordinary files you can diff, version, and audit like any other codebase .
When Zero API Keys Is Not Quite Zero
OpenMontage's "zero API keys" claim is real for the default path but does not make every capability free. Out of the box you get local, offline narration from Piper TTS, footage and stills from Archive.org, NASA, and Wikimedia Commons, word-level transcription from WhisperX, composition in Remotion, and post-production in FFmpeg — no keys required . That free real-footage path, not stitched stills, is what separates it from slideshow tools.
Generated clips are key-gated. Kling, Runway Gen-4, Google Veo 3, MiniMax, Grok Imagine, and HeyGen all require paid access, typically routed via fal.ai; local WAN 2.1, Hunyuan, and CogVideo are key-free but GPU-heavy . The README's own cost examples show the spread:
| Production | Cost | What it uses |
|---|---|---|
| 60-second animated short | $1.33 | Six Kling clips via fal.ai |
| FLUX-image animation | $0.15 | FLUX generated stills |
| 70-second history piece | $0.02 | OpenAI narration, free Pixabay music |
A default budget cap of roughly $10 keeps runaway spend in check. The bigger caveat is legal, not financial: the repo is AGPL-3.0, so its copyleft obligations attach to any distributed software incorporating the code — worth a review before shipping a product.
"Local footage stays on your machine, composition and rendering run locally without watermarks or metering, and commercial use is available on every tier including free." — OpenMontage Studio (source: daily.dev)
OpenMontage Studio — a hosted layer still in private alpha — offers commercial tiers as the cleaner path for teams uneasy about AGPL distribution .
Beyond the Animated Explainer: Eleven Other Formats
The animated explainer is one of twelve production pipelines shipped in the current repo . The named set includes animation, avatar spokesperson, cinematic, clip factory, documentary montage, hybrid, localization/dub, podcast repurpose, screen demo, and talking head — each governed by the same approval-gated, checkpoint-based loop you already ran once .
Documentary montage is the differentiating entry: instead of animating static images, it builds a CLIP-searchable corpus of real motion footage from Archive.org, NASA, and Wikimedia Commons, then cuts together actual clips — the free real-footage path no other entry-level tool offers . Localization/dub extends an existing video rather than rebuilding it; podcast repurpose turns audio-first content into a visual format with minimal new assets .
Expanding is mechanical: open the target pipeline's skill file in skills/pipelines/ and issue a fresh brief. The director loop you learned on your first render carries over unchanged to all twelve.
Frequently asked questions
Does OpenMontage actually work without any API keys?
Yes, on the default path. Narration runs on Piper (local, offline TTS), footage and stills come from Archive.org, NASA, and Wikimedia Commons, transcription uses WhisperX, composition runs on Remotion, and post-production runs on FFmpeg — all locally with zero keys . Keys only unlock generated clips (Kling, Runway Gen-4, Veo 3) and premium voices. The repo's own examples include a $0.02 run for a 70-second history piece using OpenAI narration and free Pixabay music, and free-source-only workflows are explicitly supported .
Which coding agents does OpenMontage work with?
Six directors are supported: Claude Code, Cursor, GitHub Copilot, Codex, Windsurf, and OpenCode . Configuration is file-routed per agent — Claude Code reads CLAUDE.md then AGENT_GUIDE.md, Cursor uses CURSOR.md plus .cursor/rules/, and Copilot, Codex, and Windsurf each get a dedicated platform Markdown file. All of them point to the same shared operating guidance in AGENT_GUIDE.md .
What exactly is an OpenMontage skill file?
A skill is a Markdown file encoding a reusable production capability — not a chat prompt. Examples include beat-detecting a music track, fitting an irregular image into 16:9 without stretching, or running a QA checklist . The agent reads the relevant skill files before invoking any tool. The current README advertises 700+ skill and production-knowledge files, up from the roughly 500 cited at launch .
Does the AGPL-3.0 license allow commercial use?
Yes, AGPL-3.0 permits commercial use, but with a condition: if you ship software that incorporates the code — including over a network — you must distribute your source modifications under the same license . Teams building a product on OpenMontage should audit this obligation before shipping. The hosted OpenMontage Studio, currently in private alpha with beta pending, states commercial use is available on every tier and may be the more practical path for a proprietary deployment .
How do I choose which pipeline to use for my video?
You do not choose manually. The agent reads AGENT_GUIDE.md and selects a pipeline from your plain-English brief, then presents a proposal you approve before any assets are touched . AGENT_GUIDE is prescriptive: every video request must go through a named pipeline, and ad hoc scripts or bypassing the review checkpoints are explicitly disallowed .
Top comments (0)