OpenMontage Video Studio
A coding-assistant-driven production system for planning, creating, reviewing, and rendering videos.
- Source repo
- calesthio/OpenMontage
- Stars
- ★ 77k
- Last updated
- 2d ago
- License
- AGPL-3.0
- Primary language
- Python
- FA score
- 48/100 · Major gaps
At a glance
- How it runs
- Works with
- Portable with changesCodex · Claude Code
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- An education creator making a 45-second explainer with narration, captions, and animated visuals.
- Not a fit if
- Users who want a browser-only tool and won't install Python, Node.js, and FFmpeg locally
- Teams needing fully unattended batch runs without the proposal, script, and asset approval gates
- Creators without paid API keys who still need cloud image, video, or voice generation
- Source review
- 48/100 · Major gaps
What does this agent do, and when should you use it?
OpenMontage is a local, agent-first video-production system driven by an AI coding assistant, covering research through final delivery checks. Its workflow is expressed through YAML pipeline manifests, Markdown director skills, and Python tools rather than a standalone code orchestrator. It supports Remotion, HyperFrames, and FFmpeg for composition and rendering, producing videos with generated or retrieved visuals, narration, music, and subtitles. A documentary path can assemble real motion footage from open archives and stock sources such as Archive.org, NASA, and Wikimedia Commons. It is best suited to creators or teams prepared to run Python, Node.js, and FFmpeg locally and participate in approval gates.
A user describes a video in natural language; the assistant reads a YAML manifest in pipeline_defs/ and the relevant stage-director skills in skills/, then moves through research, proposal, script, scene_plan, assets, edit, and compose. Python tools under tools/ handle retrieval, image or video generation, TTS, music, subtitles, analysis, and composition; the registry can be discovered with registry.discover() and queried through support_envelope() and provider_menu(). The system writes JSON checkpoints and keeps a decision trail for provider, style, music, voice, and renderer choices. Before composition it checks the delivery promise, slideshow risk, and renderer configuration; after rendering it runs ffprobe validation, frame extraction, audio analysis, and subtitle checks. python -m backlot open opens the local Backlot library or a project’s live production board.
- An education creator making a 45-second explainer with narration, captions, and animated visuals.
- A documentary or video-essay maker who wants to retrieve real motion footage from Archive.org, NASA, Wikimedia Commons, and optional stock sources rather than pay for a video-generation model.
- A product-marketing team producing a short launch teaser for a product concept and needing cost estimates before assets are generated.
- A podcast operator turning a long source into a ranked batch of short-form social clips.
- A training or product team creating screen-recording demos, avatar-led explainers, or localized and dubbed versions of existing video.
How do you install or deploy this agent?
Prerequisites are Python 3.10+, FFmpeg, Node.js 18+, and an AI coding assistant that can read files and execute Python. Run:
git clone https://github.com/calesthio/OpenMontage.git
cd OpenMontage
make setupThen ask the coding assistant: "Make a 60-second animated explainer about how neural networks learn". API keys are not required to start; configure only the cloud services you need in .env, such as FAL_KEY, OPENAI_API_KEY, ELEVENLABS_API_KEY, or PEXELS_API_KEY. Without make, the documented setup is to create a .venv, install requirements.txt, run npm install in remotion-composer, install piper-tts, and copy .env.example to .env.
How do you use this agent?
Give the coding assistant a brief with duration, subject, visual style, source constraints, and target platform. For real-footage production, request a “documentary montage” and explicitly say “use real footage only.” The assistant selects a pipeline and pauses for approval at the proposal, script, scene-plan, asset, and publish gates. At proposal time it selects and locks render_runtime: Remotion is the documented default for data-driven explainers, while HyperFrames fits HTML/GSAP motion-graphics work. Run python -m backlot open to view all local productions, or python -m backlot open <project-id> to inspect one production board.
What are this agent's strengths and limitations?
- Connects research, proposal, scripting, assets, editing, composition, and post-render review in explicit stages with enforced human approvals.
- Offers a real-footage documentary-montage route that builds and semantically retrieves from open and stock motion-media corpora, not only image animation.
- Uses delivery-promise checks, slideshow-risk scoring, ffprobe, frame sampling, audio analysis, and subtitle verification around rendering.
- Can start with local Piper TTS, FFmpeg, Remotion, and open-media sources while allowing optional connections to multiple cloud providers.
- Includes the local Backlot board and a persistent decision trail for production status, approvals, spend, and run replay.
- It requires a local Python, FFmpeg, and Node.js environment plus a coding assistant able to read files and execute code; it is not a browser-only video service.
- Cloud image, video, voice, and music capabilities require their respective API keys and may incur provider charges.
- The workflow intentionally pauses for approval at proposal, script, scene-plan, asset, and publish gates, which adds overhead for fully unattended batch work.
- The HyperFrames runtime requires Node.js 22 even though the general installation prerequisite is Node.js 18+, so that path has an additional version requirement.
- The README’s stated total of 12 pipelines does not fully align with the named pipeline table, so adopters should verify the required pipeline locally.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| OpenMontage Video Studio This agent | 48 · Major gaps | CLIFree + model costs | ★ 77k | 2d ago | Python | Codex · Claude Code |
| HTML Video Studio | 53 · Major gaps | CLIFree + model costs | ★ 4.8k | 4mo ago | HTML | Codex · Claude Code · Claude API |
| Video Shotcraft | Insufficient evidence | Agent plugin / skillFree | ★ 17k | 7d ago | TypeScript | Codex · Claude Code |
| ArcReel Video Creation Workspace | 52 · Major gaps | Self-hosted serviceFree + model costs | ★ 6.2k | today | Python | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows: README explicitly requires human approval for every creative decision, and Backlot storyboard acts as an approval gate, with tests verifying gate integrity (e.g., completed without approval is rejected). However, least privilege is weak: README lists many API keys (FAL, OpenAI, Google, etc.) without stating permission scope or minimization; data flow transparency only mentions Backlot showing provider decisions and spend, but not detailed data flows; sensitive data handling does not mention key storage or encryption; dependency security lacks vulnerability scanning or pinning; external effects (network requests, file writes) not clearly stated; rollback mechanism not mentioned; source attribution only mentions citing research, but not code provenance. Thus most criteria score 1, user confirmation scores 2.
Evidence shows: README describes a consistent pipeline flow (research -> proposal -> script -> scene_plan -> assets -> edit -> compose), and tests verify gate integrity, but no failure message examples or error handling documentation. Dependency availability: requirements.txt lists dependencies but no version compatibility or alternatives. Thus self-consistency scores 2, dependency availability and failure messages score 1.
Evidence shows: README targets multiple audiences (AI coding assistants, OpenClaw agents, content creators), provides many scenarios (explainers, documentaries, animations, etc.) and prompt examples. Capability boundaries are explained via pipeline list and tool registry, but trigger precision (how to precisely trigger a specific pipeline) is not detailed. Environment fit: provides macOS/Linux/Windows install instructions, but no specific GPU or cloud requirements. Thus audience and scenarios, capability boundaries, environment fit score 2, trigger precision scores 1.
Evidence shows: README provides clear information architecture (table of contents, quick start, pipelines, prompts, etc.), detailed install notes (make setup, manual commands), and rich examples and FAQ (prompt gallery, cost examples). But naming stability not mentioned (no version history or API stability); known limitations not explicitly listed; license is AGPL-3.0 but impact not explained; versioning/changelog not mentioned; maintenance responsibility not clear (only sponsorship and community). Thus information architecture, install notes, examples and FAQ score 2, others score 1.
Evidence shows: README showcases multiple finished videos (e.g., 'THE LAST BANANA' cost $1.33), output usability is high (provides final video and cost). Marginal value is clear (from zero to full video production), cost-benefit is good (low-cost examples). But no independent verification or user feedback, so score 2.
Evidence shows: README claims '700+ skill files' and '100+ tools' but no file list or verification method; tests exist but do not cover all claims; facts and inferences not clearly separated (e.g., cost examples may be conditional). Thus each criterion scores 1.
- API key handling: README requires users to add multiple API keys but does not explain key storage or least privilege, posing leakage risk.
- Dependency security: requirements.txt does not pin versions, no vulnerability scanning mentioned, supply chain risk.
- External effects: network requests, file writes, etc. not clearly stated, may exceed user expectations.
- Rollback mechanism: no rollback or recovery mechanism mentioned, failure may be unrecoverable.
- Claim verification: README claims 700+ skill files and 100+ tools but no verification method, treat with caution.