OpenMontage
github.com/calesthio/openmontage- Category
- AI Agents
- Rank
- No. 685Tools index
Previous survey · No. 1967 ·
- Pricing
- Open Source
- Type
- AGENT
- Use case
- Design & Media · Workflow Automation
- Interfaces
- Agent Skill / Plugin
- Builder
- calesthio
- GitHub
- 68.3k stars
- Date
About
An open-source agentic video production system that takes a brief plus your own materials and runs a nine-stage pipeline — research, script, scene plan, approval gate, narration, music, compose, render — to produce a finished, editable video. Costs are estimated and held at an explicit approval gate before any paid generation happens, and composition and rendering run locally with a provenance report covering source, licence, and model for each asset. It runs as a plugin inside existing coding agents including Claude Code, ChatGPT Codex, and OpenCode, as well as its own Studio app.
What it does
OpenMontage turns a coding assistant such as Claude Code or Cursor into a film crew. There is no director program inside it: the assistant reads a recipe for the chosen kind of video, follows written instructions for each step, calls Python helpers that fetch stock footage or pay a provider for voices, images and clips, and saves its progress to disk after every step. Rendering happens on your machine with FFmpeg plus Remotion or HyperFrames. A local web board shows the script, scene cards, spend and pending sign-offs as the run unfolds.
Why it's ranked here
The design choice is the argument. Because your existing assistant is the brain, the project never calls a language model API itself, so there is no second model bill. The Python side is deliberately small: helpers, saved state, and checks that refuse to advance a stage whose earlier stages are unfinished or unapproved. A dozen ready recipes cover explainers, trailers, talking heads, podcast clips, dubbing and screen demos, and every capability is meant to have both a paid cloud route and a free local one. It is AGPLv3 licensed and ships a spending policy with a per-action approval threshold out of the box.
What's good
Sign-off is enforced in code, not only requested in prose. When progress is saved, the checker looks up whether that stage needs human approval in the recipe itself, and an unknown recipe name raises an error instead of quietly skipping the check. Stages cannot be marked done while an earlier one is incomplete. Cost honesty extends to providers: for one video model with no published price, the tool demands a custom rate before quoting anything and never reports unknown pricing as free. The live board only reads project folders, never writes them, and rejects file requests that try to escape a project directory.
Tradeoffs
Most of the intelligence lives in Markdown instructions, so results depend on how faithfully your assistant follows long rulebooks; the agent guide has to spell out, in capitals, that ad hoc scripts and skipped stages are forbidden. The default spending mode warns rather than caps. Serious output needs a pile of accounts, and the provider guide lists over a dozen keys across Google, fal.ai, ElevenLabs, OpenAI, Runway and others. The installable package declares fewer dependencies than the plain requirements list, leaving out the board server libraries. AGPLv3 matters if you plan to run it inside a hosted service.
How to use it well
Best for someone already living in a coding assistant who wants repeatable short videos: explainers, product films, social clips, or a remix of a reference video you point it at. Start with the free tier the provider guide recommends, Pexels and Pixabay stock plus offline Piper narration, before adding paid keys. Switch the budget mode from warn to cap if you leave runs unattended, and keep the approval checkpoints on. Open the board to watch spend. It is not a timeline editor for hand cutting footage, and free local video generation needs a GPU.
Technical notes+
Per docs/ARCHITECTURE.md there is no runtime orchestrator: the agent reads a manifest under pipeline_defs/ (for example pipeline_defs/animated-explainer.yaml, which sets budget_default_usd 2.00, max_revisions_per_stage 3, max_wall_time_minutes 20 and human_approval_default true on proposal, script, scene_plan, assets and publish). lib/pipeline_loader.py loads manifests with yaml.safe_load, validates them against a JSON Schema, caches read-only copies via lru_cache, and exposes get_stage_human_approval_default as the single lookup shared by gate enforcement and the board. lib/checkpoint.py writes per-stage JSON checkpoints, requires a canonical artifact per stage, raises CheckpointValidationError on an unknown pipeline_type (fail closed) and, in _enforce_stage_prerequisites, blocks awaiting_human or completed writes until predecessors validate. config.yaml defaults budget.mode to warn with single_action_approval_usd 0.50. backlot/server.py is a FastAPI app that watches projects/ with watchfiles, fans changes out over SSE with bounded per-subscriber queues, and returns 403 when a media or thumb path resolves outside the project. setup.py install_requires omits fastapi, uvicorn, watchfiles and numpy, which requirements.txt lists.
Observed
- License
- AGPLv3 (README badge)
- Language
- Python tools and persistence, with a Node.js Remotion composer
- Runtime requirements
- Python 3.10+, FFmpeg, Node.js 18+, and an AI coding assistant (README.md)
- Interface
- Driven by a coding assistant reading YAML manifests and Markdown skills; local FastAPI board launched with python -m backlot
- LLM usage
- No LLM API calls at runtime; the host coding assistant is the model (docs/ARCHITECTURE.md)
- Approval gates
- Per-stage human approval read from the pipeline manifest and enforced in lib/checkpoint.py
- Budget defaults
- Warn mode, per-action approval threshold, approval required for a new paid tool (config.yaml)
- Local options
- Piper offline TTS, local diffusion images, and GPU local video generation listed in docs/PROVIDERS.md
- Packaging
- setup.py install_requires is narrower than requirements.txt and omits the board server dependencies
Read from README.md, requirements.txt, setup.py, docs/ARCHITECTURE.md, docs/PROVIDERS.md, AGENT_GUIDE.md, lib/pipeline_loader.py, lib/checkpoint.py, pipeline_defs/animated-explainer.yaml, config.yaml, backlot/server.py.
Tags
Tech Stack
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.