OpenMontage

What it is

OpenMontage is an agentic video production system: you clone the repo, open it in Claude Code, Cursor, Copilot, Windsurf, or Codex, and the assistant follows the project's instruction files to run end-to-end video workflows — research, scripting, scene planning, asset generation, editing, and final composition.

The unusual part is the architecture. There is no Python orchestrator. The repo's own docs put it plainly: "The AI agent IS the intelligence. Python exists only for tools and persistence." Orchestration, creative decisions, review, and stage transitions all live in instructions — YAML pipeline manifests plus markdown "skills" the agent reads before each stage.

How it works

Agent reads pipeline manifest (YAML)
  → reads stage director skill (Markdown)
  → uses tools (Python BaseTool subclasses)
  → self-reviews (meta skill)
  → checkpoints
  → presents to human for approval

Knowledge is organized in three layers: the tool registry declares what exists (capabilities, cost, status), project skills teach OpenMontage's conventions, and vendored tech skills explain how each underlying API actually works. Every production request runs through a pipeline state machine — idea → script → scene_plan → assets → edit → compose → publish — with enforced human approval gates at the proposal, script, asset, and publish stages.

The twelve pipelines

• Talking head
• Animated explainer
• Screen demo
• Clip factory
• Podcast repurpose
• Cinematic
• Animation
• Character animation
• Hybrid
• Avatar spokesperson
• Localization & dub
• Documentary montage

Each pipeline is a declarative YAML manifest in pipeline_defs/ defining stages, tools, review focus, and approval gates, paired with seven stage-director skills that teach the agent how to run each phase.

What it costs to run

Getting started

Prerequisites: Python 3.10+, Node.js 18+, FFmpeg, and an AI coding assistant.

git clone https://github.com/calesthio/OpenMontage.git
cd OpenMontage
make setup

Then open the folder in your assistant and ask for a video — the assistant picks up CLAUDE.md (or the Cursor/Copilot/Windsurf equivalent), which routes it through AGENT_GUIDE.md, the project's operating contract. API keys are optional and only expand the provider menu; copy .env.example to .env to add them.

Field notes from a hands-on inspection

Because this project works by having your AI assistant follow instructions from the repo, it deserved a trust check before recommending it. A clone-and-read review of the current main branch found:

• ✓Clean instruction files. CLAUDE.md and the 46 KB AGENT_GUIDE.md are a legitimate operating contract — pipeline routing rules, tool-selection discipline, and a "decision communication" requirement that the agent announce provider, model, and cost before any paid call. A scan for prompt-injection patterns, hidden exfiltration, and obfuscated payloads came up empty.
• ✓Boring, safe setup. The Makefile creates a venv and runs pip install / npm install — no piping remote scripts to a shell.
• ✓Small core dependency footprint. The base requirements.txt is about a dozen mainstream packages (Pydantic, FastAPI, Pillow, provider SDKs).
• ✓Guardrails are real. Approval gates are enforced by checkpoint tooling rather than left to the agent's discretion, and budget reservation happens before spend.