OpenMontage

OpenMontage turns an AI coding assistant such as Claude Code, Cursor, Copilot, Windsurf or Codex into a video production studio. You describe the video in plain language, and the agent works through a pipeline: research, proposal, script, scene plan, assets, edit, compose and publish. It calls a registry of Python tools for narration, image and video generation, stock footage, transcription and rendering.

It supports 12 production pipelines, including animated explainers, cinematic trailers, documentary montages from real footage, talking heads, avatar spokespeople, screen demos, clip factories, podcast repurposing, character animation and localization dubbing. You can also paste a reference video, and it analyzes pacing and structure and proposes concepts before spending anything.

It works with no API keys, using free offline narration, open footage archives, Remotion and FFmpeg. Adding provider keys unlocks more image, video, voice and music options, and a local GPU can run open video models. A local live storyboard called Backlot shows the production filling in and holds creative approvals until you answer.

What You Get

• 12 production pipelines from explainers to documentary montages
• A registry of 100+ tools across narration, image, video, audio and analysis
• Backlot, a local live storyboard with approval gates and run replay
• Budget tracking with estimates, reservations and per-action approval limits
• Two render runtimes (Remotion and HyperFrames) plus FFmpeg post-production
• Instruction files for Claude Code, Cursor, Copilot, Windsurf and Codex

Common Use Cases

• Explainer videos from a plain-language brief
• Cinematic trailers and mood-driven edits
• Documentary montages cut from free archival footage
• Social clips remade from a reference video
• Localized dubs of existing videos

Under The Hood

Architecture Agent-first orchestration: there is no Python orchestrator. The coding assistant reads a YAML pipeline manifest, follows a stage-director skill written in Markdown, calls Python tools through a registry, and writes a checkpoint for each stage. Python supplies tools and persistence, and all judgment lives in the instructions. Tools share one contract that declares capability, runtime, dependencies, cost estimation and fallbacks, and they are auto-discovered. Selector tools route between providers by ranking what is actually available. Knowledge is layered: tool contracts, project conventions, and vendored technology skills. The local Backlot server derives its board from project files and event logs, so it needs no agent involvement.

Tech Stack Python with Pydantic, JSON Schema validation, YAML manifests, FastAPI and Uvicorn for Backlot, and watchfiles for live updates. Composition uses Remotion (React and TypeScript) and HyperFrames (HTML, CSS and GSAP via npx), with FFmpeg for encoding, mixing, subtitles and grading. Piper provides offline speech, and provider clients cover fal.ai, Atlas Cloud, ElevenLabs, OpenAI, Google and others. Local models run through diffusers-style GPU paths. Setup is a Makefile with Windows and PowerShell fallbacks.

Code Quality Every canonical artifact is validated against a JSON schema, and checkpoint writes enforce stage prerequisites and human-approval gates, failing closed on gate violations or unknown pipeline names. Superseded checkpoints are archived so a run’s history can be reconstructed, and event instrumentation is deliberately non-fatal. The test suite covers contracts, tools, pipelines, styles, the board and a quality-evaluation harness, and CI runs a lint smoke check and the tests on every push and pull request. Docs include an architecture guide, a provider guide and a review guide, and the code has thorough docstrings that explain why.

What Makes It Unique It is an operating system for agents rather than an app: the agent is the control plane, and no language model API key is needed at runtime. Providers are ranked from the live registry rather than a hardcoded order, and every selection is logged. Reference-video ingestion produces a grounded plan with honest costs before spending. The system also polices its own output, checking for slideshow-style fakes and delivery promises that were not met.

Self-Hosting

OpenMontage is AGPL-3.0, a strong copyleft licence that also covers network use. You can use and modify it freely, but if you run a modified version as a service for others, you must offer them the source. Contributions are accepted under the same licence with no CLA. The README also carries sponsor placements for cloud services, which are separate products.

It is not a hosted service. You need Python 3.10 or newer, FFmpeg, Node.js 18 or newer (HyperFrames wants a newer Node), and an AI coding assistant that can read files and run code. A GPU is optional and only needed for local video generation. The free path (Piper narration, open footage, Remotion) needs no keys, but paid image, video, voice and music providers are billed by those providers, and the built-in budget controls only track spending. You also carry the setup, API keys and disk space for generated assets.

There is no licence key, paid tier or gated feature in the repo. The real costs are provider API charges and the assistant subscription you already use. Support is community-based through GitHub, with no SLA.