OpenMontage

OpenMontage is an open-source system that turns an AI coding assistant into a video production pipeline. Given a plain-language brief, an agent such as Claude Code, Cursor, GitHub Copilot, Codex or Windsurf researches the topic, writes a script, plans the scenes, generates or retrieves the images, footage, narration and music, and renders a finished video, stopping for a person's approval at each creative stage. It was created by the GitHub developer Calesthio and published on March 29, 2026 under the AGPL-3.0 license. On October 24, 2026 its repository had 71,396 GitHub stars and 9,247 forks: more than Remotion (66,968) and about level with HeyGen's HyperFrames (71,178).

How It Works

OpenMontage has no orchestration program of its own. In the project's words, "Your AI coding assistant IS the orchestrator." The repository gives the agent three layers to work from:

• Pipelines. Twelve production pipelines, each a YAML manifest: animated explainer, animation, avatar spokesperson, character animation, cinematic, clip factory, documentary montage, hybrid, localization and dubbing, podcast repurpose, screen demo and talking head. A manifest lists the stages, the skills each needs, what a review checks, where a person must approve, and default budget and time limits.
• Skills. Markdown "director" skills that tell the agent how to run each stage, plus knowledge packs about the underlying technologies; the project counts more than 700 skill and production-knowledge files (see SKILL.md).
• Tools. A self-registering Python registry of more than 100 tools for generation, audio, graphics, analysis and enhancement.

Every pipeline runs the same sequence: research (including web searches), proposal, script, scene plan, assets, edit and compose. Each stage's output is a JSON file checked against one of more than 20 schemas, from the scene plan to the edit decisions and cost log, and checkpoints can be resumed. That makes OpenMontage an example of video as code applied to a whole production.

Models, Footage and Rendering

OpenMontage's provider guide lists adapters for more than 20 cloud video-generation services, including Kling, Google's Veo, Runway, Seedance, MiniMax, OpenAI's Sora and Higgsfield, and for open models on a local GPU such as Wan, HunyuanVideo, CogVideoX and LTX-Video, some of them through ComfyUI. Other adapters cover images, voices such as ElevenLabs, and music such as Suno. These are integrations, not tested results: the repository says its tests are mocked and that "credential presence indicates configuration, not verified service access."

It can also make videos with no API keys. Piper narrates offline, and the documentary-montage pipeline builds a searchable corpus of real footage from Archive.org, NASA and Wikimedia Commons (plus Pexels and Unsplash with free developer keys), then picks clips by meaning with CLIP embeddings; see stock and archive footage. WhisperX transcription, scene detection and CLIP and BLIP-2 analyze source footage and reference videos, so a user can paste a YouTube Short and ask for something like it.

Rendering uses three engines, chosen at the proposal stage: Remotion (React) for data-driven explainers, HyperFrames (HTML, CSS and GSAP) for motion-graphics-heavy briefs and character animation, and FFmpeg for assembly, encoding, subtitles and mixing, with a pan-and-zoom fallback if Remotion isn't installed. OpenMontage's bundled Remotion skill notes that companies may need a license for commercial use; Remotion's own license requires one for companies of four or more people.

Approvals and Budgets

Approval gates at the proposal, script, scene plan, generated assets and publish stages are enforced: a gated stage cannot be marked complete without a recorded approval (see human in the loop). Each provider choice is scored on seven weighted dimensions, led by task fit (30%) and quality (20%), and logged with the alternatives. Spending is estimated first, with a $10 total budget by default and an approval for any action above $0.50. After rendering, a self-review samples frames and checks audio levels and subtitles. A local web board called Backlot shows stages, assets and spending live and holds the storyboard for approval. The project says one 60-second animated showcase cost $1.33 to generate.

Status as of October 2026

• No releases yet. The repository has no tagged releases, setup.py says version 0.1.0, and development happens on the main branch (497 commits by October 24, 2026).
• Community. 73 contributors, with Calesthio the author of 318 commits, and 298 pull requests open against 152 merged. Coverage so far has come from developer blogs and YouTube rather than the technology press.
• Setup. git clone and make setup, with Python 3.10+, FFmpeg and Node.js 18+ (22+ for HyperFrames), on macOS, Linux or Windows.
• Gaps. No MCP server and no timeline export to traditional editors were found (a code search for FCPXML and OpenTimelineIO returns nothing); output is an MP4 plus a local export bundle. Local language models through Ollama and LM Studio are listed as coming soon.
• License. AGPL-3.0 is a copyleft license: anyone who distributes a modified version, or lets people use one over a network, must make its source code available under the same license.
• Studio. The project's website, openmontage.video, describes OpenMontage Studio, a desktop app in private alpha (with a waitlist) that can use the user's own provider keys or "OpenMontage credits".

How It Compares

OpenMontage works further upstream than most AI editors. Descript and VEED are shipping editors built around recorded footage (see Descript vs OpenMontage), and VEED's OpenEdit is a smaller Apache-2.0 agent skill that writes edits as HTML pages (see OpenMontage vs VEED OpenEdit). Its strengths are openness, breadth of models and footage, and a process anyone can read and change; its limits are a young codebase, no hand-off to editors such as Premiere or DaVinci Resolve, and results that depend on the agent and models a user configures. See also agentic video editing and AI filmmaking.