An AI production studio for full YouTube videos
OpenMontage from an empty laptop: a paid Claude plan, four free helper programs, five setup steps on Windows or Mac, seven copy-paste prompts for full videos, Shorts, ads and documentaries, the pro moves, the mistakes that cost money, and a first-week plan. Checked against the project’s own code, Anthropic’s docs, the providers’ pricing and YouTube’s policies, with the steps that no longer work fixed
OpenMontage turns an AI coding assistant into a whole video crew: it researches, writes the script, plans the scenes, finds real footage or generates visuals, narrates, adds music and captions, edits and renders an MP4, and stops for your OK at every creative decision. This is the beginner’s route in, from a laptop that has never opened a terminal to a first finished video, with every command and all seven prompts. The checks turned up a handful of steps that no longer work as written. Claude Code now starts in auto mode, so it doesn’t ask before each command unless you switch it; on a Mac, Homebrew’s Python 3.12 isn’t the python3 the setup finds; Piper installs without a voice; new Pexels keys are paused; the budget limits are rules the agent follows, not locks in the code; and the free ElevenLabs tier can’t be used in anything you monetize. Each fix is marked where it applies.
OpenMontage is a free, open-source project that turns your AI coding assistant (this page uses Claude Code, Anthropic’s assistant that works on your computer; Cursor, Copilot, Codex and Windsurf work too) into a complete video production team. You describe the video in plain English. It researches the topic, writes the script, plans every scene, finds footage or generates visuals, records the voiceover, adds music and captions, edits it all together and hands you a finished MP4. This is the beginner’s route in; the Lab’s earlier OpenMontage entry covers the same project from the developer’s side.
What held up, on 8 October 2026
Checked against the OpenMontage repository at commit 9327439 (3 October 2026), read rather than run; Anthropic’s Claude Code docs and pricing; the winget and Homebrew package indexes; Pexels, Pixabay, Google Cloud, ElevenLabs and fal.ai’s own pages; and YouTube’s help centre.
A production crew, not a clip generator
It is not a side project nobody uses. The code lives on GitHub (think of GitHub as Google Drive for software), where it had about 65,000 stars, GitHub’s version of likes, on 8 October 2026. It reached #1 Repository of the Day on GitHub Trending in June 2026, and was last updated on 3 October.
The analogy that makes it click: picture Claude Code as a very capable new hire who can use your computer. OpenMontage is the studio you hand them on day one: a thick handbook on how a real production crew works (700+ skill and knowledge files) and a fully stocked equipment room (100+ production tools).
The crew you get, and what each one does
The researcher
Runs live web searches (YouTube, Reddit, Hacker News, news, studies), typically 15 to 20 and never more than 25, before a single word is written, so the facts are current, not made up. (The guide says 15 to 25+; the project’s research playbooks cap it at 25.) This happens in the explainer, animation and cinematic pipelines; montages and clip jobs skip research.
The scriptwriter
Writes the narration with voice direction, then waits for your OK.
The storyboard artist
Breaks the script into scenes and shows you a scene plan before anything is generated. (The guide calls it a visual plan; it is a written one, and the images come at the next stop.)
The footage hunter
Pulls real motion clips from free sources like Pexels, Archive.org, NASA and Wikimedia Commons. Archive.org, NASA and Wikimedia need no key at all.
The visual artist
Generates AI images and AI video clips if you connect a provider. Optional, and pay-per-use.
The voice
A free, offline narrator voice, Piper, comes with setup. Premium voices plug in if you want them.
The editor
Cuts the timeline, mixes music, burns in word-by-word captions and renders the final file.
The quality checker
Checks its own render before it shows you the video: sampled frames for black frames, the audio for silence, and whether captions are there. A silent or broken file stops it outright; a black frame or missing captions gets flagged for you. (The guide says it watches its own render; it samples it.)
Most AI video tools give you one clip of a few seconds (the guide says 5 to 10) from a prompt. A real YouTube video is fifty decisions strung together: topic, angle, script, shots, voice, music, pacing, captions, export. OpenMontage is the thing that strings them together, in order, like a real production team would.
Six words you will see on this page
There is no OpenMontage app
Most people expect a website with a big Generate button. There isn’t one. Your AI assistant is the producer: it reads the playbook, picks the tools and runs them. You talk to it like you would talk to a producer you hired. (A separate desktop app, OpenMontage Studio, is in private alpha behind a waitlist. The open-source project on this page has none.)
That sounds harder, but it is the reason it works. A button can only do one thing. A producer can research, ask you questions, change plans and fix its own mistakes.
OpenMontage is built to stop and wait for you at every creative decision: the proposal, the script, the scene plan, the generated visuals and the final publish. It shows you what it made and what it will cost, then it ends its turn until you say yes. Nothing gets made behind your back. The exact stops vary a little by pipeline; a real-footage montage, which has no script, stops at the idea, the scene plan, the footage and the edit instead.
Nothing to start
With zero API keys you already get a free narrator voice (Piper), free real footage from open archives, animated charts and text scenes, captions and a full edit. The only thing you pay for is your Claude plan, and your videos draw on the same usage allowance as your Claude chats. (One exception: Remotion, one of the two animation engines, needs a paid company licence if you run or work for a for-profit company of more than three people; a client’s size doesn’t count when you only hand them the finished video. See “Mind the licences” below.)
When you do add paid AI visuals, the costs are small. The project’s own published examples:
It estimates the cost before it spends anything, names the tool, provider and model before any paid call, shows the running cost at every approval stop, and gives each video a budget of $0.50 to $3 depending on the pipeline. The project’s settings also list a $10 overall cap, approval for any single step over $0.50 and approval before the first use of any paid tool. Treat all of that as rules the agent follows, not locks in the code: nothing in the project checks the budget while it runs, so the $10 cap only warns on paper, and switching its mode to cap in config.yaml doesn’t create a hard stop either. Prompt 6 gives the agent a firmer number for each video, but it is an instruction too. The firmest limits are the ones you set at each provider (see “Four more habits” below): a prepaid fal.ai balance, and ElevenLabs’ per-key quota and monthly cap. Google Cloud budgets only alert.
This is not a one-click toy. Plan about an hour for the first install and 10 to 20 minutes of production time per short video after that, plus the time you spend reading at each stop; real-footage montages and AI-video pieces can take 30 to 60 minutes. The free path gives you clean explainer and documentary-style videos, not Hollywood. Always watch the final video yourself before you post it.
1. A paid Claude plan, for Claude Code
This is the part most guides skip, and it is why most people get stuck. Read it once and the rest is copy-paste.
Claude Code is Anthropic’s AI assistant that works directly on your computer. Unlike the chat website, it can open files, install programs and run tools.
It is included in every paid Claude plan. Pro is $20 a month ($17 a month billed yearly, $200 up front). Max starts at $100 a month for five times Pro’s usage, or $200 for twenty times, which helps if you plan to make videos every day. The free Claude plan does not include Claude Code. Subscribe at claude.com/pricing before you start. Those are US prices, before tax.
2. A computer that can run it
3. Terminal basics (2 minutes, promise)
Open it on Windows
Press the Start button, type PowerShell, press Enter. Pick Windows PowerShell, not the one marked (x86). You will see a line that starts with PS C:\Users\YourName. On Windows 11 it may open inside an app called Terminal, which is fine as long as the line starts with PS.
Open it on Mac
Press Cmd + Space, type Terminal, press Enter.
Paste a command
Copy it from this page, then right-click inside PowerShell or press Ctrl + V (Windows), or press Cmd + V (Mac). Press Enter to run it.
Wait for the prompt
A command is finished when the blinking cursor comes back on a fresh line. Do not type the next one before that.
“cd” means “go into a folder”
cd OpenMontage = step inside the OpenMontage folder.
After installing anything, close the window and open a new one
Otherwise the terminal does not notice the new program yet.
Red text is not always a disaster
Copy the full message and paste it into Claude Code. It will explain it in plain English.
4. Four free helper programs
OpenMontage leans on four free, widely used programs. You install each one once and forget about it.
Install Claude Code
Open your terminal and paste the line for your computer. It downloads Anthropic’s official installer and sets Claude Code up. On Windows, in PowerShell (it shows no progress while it downloads; wait for “Claude Code successfully installed!”, and if it says irm is not recognized you are in the older Command Prompt, not PowerShell):
On a Mac, in Terminal (the same line works on Linux):
Check it worked
When it finishes, close the terminal, open a new one and run this. You should see a version number, something like 2.1.286 (Claude Code).
Now type claude and press Enter. A browser window opens. Log in with your paid Claude account and come back to the terminal. If no browser opens, press c to copy the login link and paste it into one; when the terminal says Login successful, press Enter.
Two alternatives if you already use a package manager: winget install Anthropic.ClaudeCode on Windows, or brew install --cask claude-code on a Mac. Neither updates itself the way the line above does, so you would update by hand.
The Claude desktop app (claude.com/download) includes Claude Code too: open its Code tab and you can work with it in a normal window, on a paid plan. Step 5’s start-up lines don’t run there, so begin each session by telling Claude Code: “Activate the project’s .venv before every command (on Windows, put .venv\Scripts first on PATH), so the project’s Python and the Piper voice are found.” For your first week, also pick Manual in the mode selector next to the send button (Shift + Tab doesn’t switch modes in the app). This page uses the terminal because every step is a single copy-paste you can follow exactly.
Newer versions of Claude Code start in auto mode, where a built-in safety checker approves routine steps without asking you. That is convenient later, but for your first week you want to see every command before it runs. Start it with claude --permission-mode manual, or inside Claude Code press Shift + Tab until the bar at the bottom says manual mode on. Then it asks before most commands and file changes. Shift + Tab only lasts for that session: every new claude starts in auto mode again. So for week one start it as claude --permission-mode manual every time (the Step 4 and Step 5 lines below do), or ask Claude Code to make Manual its default mode in your settings. Drop the flag once you are comfortable. (The guide describes Claude Code as asking before it does anything. Manual mode comes closest: it asks before most commands and file changes, though it still reads files without asking.)
Install the four helpers
The easy way, and the recommended one: let Claude Code do it for you. On an Apple Silicon Mac that doesn’t have Homebrew yet, first install Homebrew yourself in Terminal, as “The manual way on Mac” below shows, including the “Next steps” lines it prints: it needs your password, which Claude Code can’t type for you. On an Intel Mac, skip Homebrew and add “This is an Intel Mac, so don’t use Homebrew.” to the end of the prompt. Then open a new terminal, type claude --permission-mode manual, and paste this.
Why it helps: Claude Code checks what you already have, skips what is installed and fixes small problems on the spot. What you get back: a short checklist of what was missing, then a permission request for each install. Approve them one by one with plain Yes.
The manual way on Windows
Paste these one at a time into PowerShell, waiting for each to finish. winget is the app installer that comes with Windows 11 and recent Windows 10; if PowerShell says winget is not recognized, update App Installer from the Microsoft Store. The first time, winget may ask you to agree to its source terms: type Y and press Enter. Click Yes on any Windows permission pop-up.
The manual way on Mac
First install Homebrew, the Mac app installer for developer tools. Paste this line, press Enter and follow its prompts. It asks for your Mac password, and nothing shows on screen while you type it, which is normal; you need an administrator account. Press Enter when it says Press RETURN/ENTER to continue. It may also install Apple’s command line developer tools, which takes a few minutes; if a pop-up asks to install them, click Install. When it finishes it prints a “Next steps” list: copy and run those lines too.
Then install the four helpers in one go. One catch: Homebrew’s standard ffmpeg is now a slimmed-down build without the filter that burns subtitle files into a video. OpenMontage’s word-by-word captions are drawn by its animation engine instead, so you may never notice; if a render fails with “No such filter: subtitles”, ask Claude Code to install ffmpeg-full and put it on your PATH.
Check all four
Close the terminal, open a new one and run these. Each line should print a version number. On Windows, if python is not found, try py --version. On a Mac, if you used the manual Mac lines, type python3.12 --version instead of python: Homebrew installs Python 3.12 under that name, and plain python3 there is Apple’s older 3.9, which OpenMontage won’t accept. If Claude Code installed Python for you, ask it which version and use that number, for example python3.13 --version.
Download OpenMontage
This copies the whole project (100 MB or more) into a folder called OpenMontage inside your user folder, then steps inside it. The ~ sign is shorthand for your user folder (C:\Users\YourName on Windows).
Run the one-time setup
Setup installs the Python tools, the animation engine and Piper, the free narrator voice engine (the voice itself is one extra line, below), and creates a settings file called .env where API keys go later. It also creates a virtual environment: a private box for OpenMontage’s Python tools so they never clash with anything else on your computer.
Easiest route: start Claude Code inside the folder by typing claude --permission-mode manual, and let it handle it.
The easy way
With Claude Code running inside the OpenMontage folder, paste this. Keep Windows or Mac in the brackets, whichever you are on. Changed from the guide: it also asks for the voice itself, because setup installs Piper without one and Piper no longer downloads it on first use, and on a Mac it names the Python version, for the reason given under the Mac line below.
Doing it yourself on a Mac
Run from inside the OpenMontage folder. Changed from the guide, which gives plain make setup: the setup looks for Python 3.10 and then for python3, which on a Mac is Apple’s 3.9, so it stops with “OpenMontage requires Python 3.10+”. Naming 3.12 makes it use the Python 3.12 that the manual Mac line installs; if Claude Code installed Python for you in Step 2, ask it which version and put that number in its place, for example PYTHON_VERSION=3.13, and check it with python3.13 --version. The second line downloads the voice.
Doing it yourself on Windows
Paste these into PowerShell from inside the OpenMontage folder, one at a time. They are the project README’s Windows line, adapted so nothing trips PowerShell’s script blocking: the README activates the virtual environment with a script, and these call its Python directly instead. The last line, added here, downloads the voice. The first render fetches the second animation engine, HyperFrames, by itself.
If you see “running scripts is disabled on this system”, PowerShell blocked a script: usually you typed npm or npx instead of npm.cmd or npx.cmd. To lift the block for your account once and for all, run Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned, then type Y and press Enter when it asks (pressing Enter alone means No). If npm install fails with ERR_INVALID_ARG_TYPE, the project’s own fix is npx --yes npm install, run inside the remotion-composer folder; in PowerShell type it as npx.cmd --yes npm install.
Take it for a test drive
Always start Claude Code from inside the OpenMontage folder: that is how it finds the playbook. Every time you come back, open a new terminal and type this. Changed from the guide, which has just cd OpenMontage and claude: the middle line switches on the project’s virtual environment, without which the voice tool isn’t found and the project’s Python isn’t the one that runs, and the last line keeps week one in Manual mode (drop --permission-mode manual once you are comfortable). On Windows:
On a Mac:
Now ask for the free demo renders.
What you get back: three short, silent demo videos of about 20 to 25 seconds each, rendered with no API keys at all: world-in-numbers (animated stats and charts), code-to-screen (a developer explainer) and focusflow-pitch (a startup pitch). They are saved in projects/demos/renders inside the OpenMontage folder; on a Mac, make demo renders them directly. If those play, the animation engine works. Your first real video is what tests the voice and the captions. (The guide says that if they play, your studio works.)
Updating it later
Git doesn’t update anything by itself. To get the newest version, run this, then start Claude Code and ask it to install any new requirements and keep your .env. On a Mac, make setup PYTHON_VERSION=3.12 does the same and leaves .env alone. Don’t re-run the Windows lines from Step 4 as they are: their Copy-Item line would overwrite .env and every key in it. If git pull stops with “Your local changes to the following files would be overwritten”, paste the message into Claude Code and ask it to save your change, finish the update and put the change back.
Read every permission prompt
In Manual mode, Claude Code asks before it runs a command or changes a file. In week one, answer each one with plain Yes instead of picking “Yes, and don’t ask again for…” or “Yes, and switch to auto mode”. You will learn what normal looks like very fast.
API keys are passwords
They live in the .env file inside the OpenMontage folder and nowhere else. Never post a key in a screenshot, a public chat or a comment, and never paste one into the Claude Code chat; if you want, ask Claude Code to add a rule that stops it reading .env at all. To add a key, open the file from inside the OpenMontage folder, after Step 4 has created it. On Windows:
On a Mac:
Find the line for the service (if it is not there, add it on its own line), paste your key right after the = sign with no spaces, and save. These are the names OpenMontage uses for the free and cheap services on this page. Added to the guide’s list: GOOGLE_TTS_API_KEY, which isn’t in the file yet but is the one OpenMontage reads first for Google voices (see the free keys below).
Then tell Claude Code this. It runs the project’s check and confirms what each key unlocked. “Available” means the key is filled in; the first real use is what proves it works.
Four more habits
Cap your spending at the provider too. OpenMontage has its own budget, but use the provider’s controls as well. ElevenLabs lets you set a credit quota on each key and a monthly spend cap. fal.ai has no cap for API keys: it is prepaid, so load only what you are willing to spend and don’t turn on any automatic top-up. Google Cloud budgets only send alerts for Text-to-Speech, so set one.
Know where your footage comes from. Pexels and Pixabay footage is free for commercial use with no credit required, but don’t use it to imply that the people or brands in it endorse you, or show identifiable people in a bad light. Archive.org and Wikimedia clips each have their own licence: CC BY clips need a credit, and non-commercial ones can’t go in a monetized video. NASA footage should be credited, and NASA’s logo isn’t free to use. Ask the agent to list every clip source and licence at the end of each video, and check them yourself.
Follow YouTube’s AI rules. When a video contains realistic AI-generated scenes or people, or AI music that is a main part of the video, YouTube asks you to choose Yes under “AI use” in YouTube Studio during upload, and it adds a label. Disclosing doesn’t limit a video’s reach or its ability to earn. Animated, clearly fantastical or script-help-only content does not need it, though YouTube may still label it.
Mind the licences inside the video. Remotion, one of the two animation engines, is free for individuals, non-profits and companies of up to three people, including commercial work; bigger companies need a paid Remotion company licence. The free ElevenLabs tier can’t be used commercially at all. And Piper’s default voice, en_US-lessac-medium, was trained on a dataset licensed for research only. Whether that binds the voice’s output is unsettled, but for monetized videos the careful choice is a voice with public-domain data, such as en_US-ljspeech-high: download it the same way and ask Claude Code to use it.
YouTube does not pay for “inauthentic content”: mass-produced, generic or repetitive videos made only for views. The rule, renamed from “repetitious content” in July 2025, names AI-generated videos made from generic templates, slideshows and scrolling text with little narrative, and AI personas posing as experts on health, money, law or politics. Pumping out 30 near-identical AI videos a week is the fastest way to get demonetized. Use OpenMontage to make better videos faster, not to spam.
One more: OpenMontage itself is licensed under AGPLv3. That license is about the software code, not the videos you make with it. If you change the code and offer it to other people as a service, read the license first.
Copy-paste prompts
Paste these into Claude Code while it is running inside the OpenMontage folder. Replace everything in [square brackets] with your own details.
Prompt 1: Make a full YouTube video from a channel idea
Why it works: it gives the producer what a real producer would ask for: who the channel is for, what the video must achieve, the format and the money rules. What you get back: a research brief with sources, 3 angles to pick from (sometimes after a quick mood board to check the direction), a script to approve, a scene plan, then a finished MP4, and it tells you the exact file location: projects/<video-name>/renders/final.mp4.
Prompt 2: Remix a video you already love
Why it works: starting from a reference is faster than inventing the perfect prompt, and OpenMontage is built for it. What you get back: what it will keep, what it will change, what it will cost and a 10 to 15 second sample before full production. To read a link it needs a downloader, yt-dlp, that setup doesn’t install, plus a small program called Deno for YouTube links; Claude Code will offer to install them, or ask it to.
Prompt 3: A real-footage mini documentary for $0
Why it works: the phrases “documentary montage” and “use real footage only” switch it to the pipeline that cuts real motion clips instead of animating still images (a beta pipeline). What you get back: an edited video made of actual footage from Archive.org, NASA, Wikimedia and stock libraries, plus a credits list. On a channel you plan to monetize, choose narration yes or add your own commentary: YouTube treats clips edited together with little or no narrative as reused content.
Prompt 4: Turn one long video into a week of Shorts
Why it works: it names the exact pipeline, the number of clips, the length and the format, so there is nothing to guess. What you get back: 5 captioned vertical clips, ranked, ready to schedule. Clip Factory is a beta pipeline and needs a free transcriber, faster-whisper, that setup doesn’t install; Claude Code will add it when asked.
On Windows 11, right-click the video file and choose Copy as path, then paste it into the prompt; on Windows 10, hold Shift while you right-click to see it. On a Mac, drag the file straight into the terminal window and its path appears.
Prompt 5: An explainer ad for your business
Why it works: your real numbers become animated stat reveals, which look expensive and need no AI image budget. What you get back: a narrated, captioned explainer you can post or run as an ad.
Prompt 6: The budget guardrail (add it to any prompt)
Why it works: it sets the cap in plain words, on top of the project’s own budget settings (the guide calls them a built-in $10 limit; see “No surprise bills” for what they really do), and puts one number in front of the agent for this video. It is still an instruction, not a lock. What you get back: a cost line at every approval stop.
Prompt 7: When something breaks
Why it works: it turns a scary red error into a one-paragraph explanation and keeps the agent from quietly changing your plan. What you get back: the cause and one proposed fix to approve.
Watch the production live
OpenMontage has a built-in visual board called Backlot that opens in your browser. Stages light up, the script appears as a screenplay page, every scene shows up as a card with its cost, and every provider choice and dollar spent is on the wall. The agent opens it by itself when a production starts; to open it again, ask Claude Code. When a run finishes, you can replay the whole production from start to end. The board only shows; you still approve in the chat.
Or open it yourself: open a second terminal and run only the first two lines of Step 5 (not claude), then one of these. The first opens the library of every project on disk. The second opens one production’s live board: replace YOUR-PROJECT-ID with that production’s folder name inside the projects folder.
No production yet? The project ships a simulated one, which needs one package setup doesn’t install, so the first line adds it (once). The second line runs a fake production for about a minute; the third opens its board, where ▶ Replay run plays it back. Run the third line from another terminal started the same way to watch it live instead. This isn’t in the guide.
Before rendering, Backlot shows a scene-by-scene contact sheet with each image, the model that made it, its cost and a quality score (the guide says its prompt; the prompt is recorded, but the board doesn’t show it). Fixing a weak scene here takes seconds. Fixing it after the render means redoing the edit.
Add the free keys in this order
Each one unlocks more tools. The first three cost nothing to start, with two catches: Google’s Cloud voices need a card on file and charge automatically past the free allowance, and free ElevenLabs audio can’t go in anything you monetize. The fourth is pay as you go.
Pixabay and Pexels (free, no paid tier)
Millions of free stock videos and photos (the guide says free forever; neither site promises that). For Pixabay, log in and your key appears on pixabay.com/api/docs. Pexels has paused new API keys, as its pexels.com/api page says; add one if you already have it or the pause has lifted. Archive.org, NASA and Wikimedia footage needs no key at all.
Google (free tiers)
Two different things. A Google AI Studio key (aistudio.google.com/apikey, Create API key) unlocks Google’s Gemini tools, including free Gemini text-to-speech with 30 curated voices plus hundreds more, rate-limited, and on the free tier Google may use your text to improve its products. That key goes in GOOGLE_API_KEY. Google Cloud Text-to-Speech, the voices the guide means (Google’s list now runs to about 2,000), is separate: it needs a billing account (card) on a Google Cloud project, the API switched on (the “Enable Cloud Text-to-Speech” link at the top of this page, then click Enable), and a key made under APIs & Services → Credentials and restricted to that API, saved as GOOGLE_TTS_API_KEY; new AI Studio keys won’t work for it. Free each month: 1 million characters of Chirp 3 HD, OpenMontage’s default Google voice, 1 million each of Neural2 and Studio, and 4 million of Standard and WaveNet. Beyond that it charges automatically.
ElevenLabs (free tier)
Premium voices. Sign up, then copy your key from elevenlabs.io/app/developers/api-keys. The free tier is 10,000 credits a month, about 10 minutes of narration on the default model (the guide says 2 to 3 minutes), with ElevenLabs’ default voices only; music, sound effects and the voice library need a paid plan or pay-as-you-go credit when used through a key, as OpenMontage does (the guide includes music and sound effects in the free tier). Free-tier audio must be credited with elevenlabs.io or 11.ai in the video title, and can’t be used commercially at all: for a monetized channel or an ad you need a paid plan (Starter, $6 a month, includes a commercial licence).
fal.ai (pay as you go)
Only when you want AI-generated images and video. Sign up, buy a little prepaid credit, and copy your key from fal.ai/dashboard/keys. Images cost from a fraction of a cent to about $0.15 each depending on the model and size (the guide says around $0.03). OpenMontage’s default image model, FLUX Pro 1.1, is billed $0.04 per megapixel rounded up: about $0.04 for a square image and $0.08 or more for a full-HD 16:9 frame. No monthly fee.
Ask for the pipeline by name
It picks one automatically, but naming it gets you exactly what you want faster. Beta pipelines work but haven’t been fully audited; the agent will say so when it picks one.
Name the platform
These are built-in export profiles, so the framing is right the first time. The profiles cap Shorts at 60 seconds and Reels at 90; if you want longer, say the length.
Make every episode match
Once you have a video you love, ask this. Consistency is what makes a channel feel like a brand. What gets saved is the look, the voice style and the music mood, as a custom style playbook in the styles/custom folder, next to the five that ship with the project; the playbook has no slot for structure, so restate that in each prompt.
Ask why it chose what it chose
Every provider, voice and music choice is logged with the options it weighed and the reasons; Backlot shows the same log in its Decisions panel. Ask “Show me the decision log for this video” and you will learn more about video production in ten minutes than in a week of tutorials.
If your computer has a strong NVIDIA graphics card, OpenMontage can generate AI video clips locally for free with open models like WAN and Hunyuan. Ask Claude Code to set up local video generation. The project command is make install-gpu, on Linux or a Mac, where make exists; afterwards fill in two lines that are already in .env but empty, VIDEO_GEN_LOCAL_ENABLED=true and VIDEO_GEN_LOCAL_MODEL=wan2.2-ti2v-5b (adding second copies lower down does nothing, because OpenMontage reads only the first; wan2.2-ti2v-5b is the default, about 12 GB of graphics memory; wan2.1-1.3b needs about 8 GB, hunyuan-1.5 about 14 GB). On Windows there is no make, and the ordinary install gives a PyTorch that can’t use the card, so ask Claude Code for the CUDA build. Apple Silicon Macs can run the same models more slowly. All of it is marked experimental.
And how to avoid them
Do not raise the budget cap “just to make it work” when something fails. A failing step usually needs a fix, not more spend. Use Prompt 7 first.
One step a day
Day 1: Install
Get a paid Claude plan, install Claude Code and the four helpers. Check every version number prints.
Day 2: Set up and test
Download OpenMontage, run setup (including the voice line), render the demo videos. If they play, the animation side works; Day 3’s first video tests the voice and captions. (The guide says that if they play, you are in business.)
Day 3: First free video
Make a 45 to 60 second animated explainer about something in your niche. Free tools only. Approve each stage slowly and read what it shows you.
Day 4: Add free footage
Add your Pixabay key to .env, and a Pexels key if you can get one. Make a real-footage montage with Prompt 3.
Day 5: Remix a winner
Pick the best video in your niche and run Prompt 2. Choose the strongest of the 3 concepts.
Day 6: The full YouTube video
Run Prompt 1 with your real channel idea. Open the Backlot board and approve the storyboard scene by scene.
Day 7: Lock your style
Save your best video as a style playbook, cut it into Shorts with Prompt 4, and plan your next 3 episodes. If you want AI visuals, load about $5 of fal.ai credit and tell OpenMontage the budget is $5.
In seven days you go from never having opened a terminal to running your own AI production studio: research, script, visuals, voice, edit and export, with you approving every step. That is work that usually takes a whole team. (The guide says a team that costs thousands a month to hire; nothing backs that figure.)
Built from a beginner’s guide to OpenMontage, kept in full and in its order, with its community sign-up left out. The table at the top lists what the checks corrected, and the main fixes are also marked where they apply. Checked on 8 October 2026 by reading the repository at commit 9327439 (3 October 2026), including its Makefile, config, pipeline manifests, tools and Backlot board, and the piper-tts 1.8.0 package, without running any of it; and against Anthropic’s Claude Code docs and pricing page, the winget and Homebrew package indexes, and the providers’ and YouTube’s own pages. The Backlot screenshots come from the project’s README. Prices, free tiers and Pexels’ key pause change often; check them again before you rely on them.