The MakeRoll UGC pipeline kit runs on your own Mac. It takes a short 2D animated show (script beats, characters, their voices, a visual style) through eight stages and ends with a captioned 1080 × 1920 video. It won't call the video done until its own scorer rates it 95/100 or higher.
Re-running picks up from cache. A paid image or voice line is bought again only when its inputs change, so fixing one line of the script doesn't re-buy the whole cast.
New to this? Start with free Basics, 7 short lessons in Telegram, one a day, no card.
# one command, eight stages
$ ./ugc run
1 teardown 2 script
3 cast 4 voice
5 mouths 6 animate
7 assemble 8 score
# exit 0 only at 95/100 or higher
It suits creators and small brands who want to make episodic 2D character shorts and don't mind typing commands into a terminal on a Mac. You won't write code. You edit plain text and JSON files, then run the commands on this page. It isn't a phone app, it doesn't post anything for you, and it can't promise anyone will watch.
The kit is the animated-story version of the pipeline. The photoreal UGC ads in MakeRoll's case studies came from a separate ad workflow that isn't part of the kit.
Below is the kit's starter episode as ./ugc run rendered it on 1 Oct 2026, from a fresh unzip of the kit. Nothing was edited after the score stage passed it. The copy here is re-encoded to 720 × 1280 so it loads quickly. The kit writes 1080 × 1920.
| Machine | Score | Voice spend | Wall time |
|---|---|---|---|
| macOS 26.6, Apple silicon (the video above) | 96.89 | $0.0725 | 237 s in 2 runs |
| Ubuntu 24.04 arm64, fresh container | 98.26 | $0.0725 | 253 s in 1 run |
Both runs were on 1 Oct 2026. Scores come from each run's final/quality_report.json, spend is the treg balance before the run minus the balance after, and wall time comes from the kit's own run log. A first episode with a new cast also buys its stills and mouth shapes, see running costs.
Nadia and Greg are AI-generated fictional characters. AI models made their faces and voices, and they aren't real people, MakeRoll members or customers. The video says "Made with AI" in the picture.
After a Lifetime payment the bot sends you a member key and one line for your AI agent. Paste that line into an agent that can run commands in your terminal, such as Claude Code or Codex, and it installs the kit and checks it. Nothing to unzip.
The bot's message has it with your own key filled in, so copy it from Telegram. It looks like this:
Set up the MakeRoll UGC pipeline: https://makeroll.com/ugc-pipeline/agent.md key: mrk_...The agent reads agent.md, a short public instruction file, then runs the installer below with your key.
Same result, no agent needed. Replace YOUR_KEY with the key from the bot's message.
curl -fsSL https://makeroll.com/ugc-pipeline/install.sh | sh -s -- YOUR_KEYIt downloads every kit file with your key, checks each one against its SHA-256 checksum, unpacks the kit into ~/makeroll-ugc-pipeline, saves the key to ~/.makeroll/key where only you can read it, runs the kit's own install.sh, and adds the MakeRoll skill for Claude Code and other agents. It's safe to run again, it picks up where it stopped, and it never uses sudo without asking.
It installs and verifies the kit, runs ./ugc doctor and then a free dry run, and shows you the estimated spend. It starts a paid run only after you say yes, and it never passes --force or --allow-orphans without your OK.
Sign in to treg.to, which pays for the AI calls. Run treg login, finish the sign-in in the browser that opens, and decide whether to top up. Step 2 shows how. Your agent can't sign in for you, and it shouldn't try.
Your key is personal. Don't paste it into a public chat or commit it to a repository. Lost it, or think someone else has it? Send /pipeline newkey to the bot and the old key stops working. Prefer to do it all by hand? The setup steps below cover the same ground.
The guide, the system requirements, the licence and the refund policy are public. The kit goes only to Lifetime members. The bot sends your member key, and the key unlocks the download, so there's no account to manage.
Open @MakeRollMembershipBot and pick Lifetime. Also included with the $199 one-time plan, and separate from the core service, is a personal licence to the kit. Weekly membership doesn't include it. Read the licence and the refund policy before you pay. The kit is final once it's delivered.
Open the bot →The bot replies with a secure Stripe Checkout link. Stripe handles the card, so MakeRoll never sees the number.
After verified payment the bot sends your members-room invite, then your member key (it starts with mrk_), the line for your AI agent, the installer command, and one file, the core kit zip, for a manual setup. The installer downloads the rest with your key. Lost the key? Send /pipeline newkey to the bot while your Lifetime membership is active.
Give your AI agent the line from the bot, or run the installer yourself, as shown above. The setup steps below cover the manual route and everything after the install.
Every command below has a copy button. Run them in a terminal, in order. If your agent or the installer has already set the kit up, check step 2 and go on from step 4. Steps 1 to 5 cost nothing (the top-up in step 2 adds credit, it doesn't spend it), and step 6 is the first one that can spend from your treg.to balance.
You need a Mac with macOS 13 or later, Python 3.9 or newer, ffmpeg with ffprobe, the treg command line tool 0.22 or newer (step 2 installs it), a treg.to account with credit and an internet connection. Leave about 1 GB of free disk. A clean test install with one finished episode used about 0.7 GB, plus about 0.1 GB for treg.
install.sh also has an apt route for Ubuntu 22.04 or newer, but so far only the macOS route has been tested end to end. On Windows, the only route is WSL2 with Ubuntu, and it's untested. Check Python and ffmpeg now.
python3 --version
ffmpeg -version
ffprobe -versionMissing ffmpeg? With Homebrew installed, one command adds both ffmpeg and ffprobe.
brew install ffmpegEvery paid call (character stills, voice lines, mouth shapes) goes through treg.to, which keeps the AI vendors' keys on its own servers and bills each call to a prepaid balance you control. You pay treg directly, per call, and only for calls the pipeline makes. You don't need accounts with Google, OpenAI or MiniMax.
The treg tool needs Python 3.12 or 3.13, and a stock Mac ships 3.9, so its installer stops unless uv is there first. uv downloads the right Python by itself and doesn't touch your system Python. Skip this if uv --version already prints a version.
curl -LsSf https://astral.sh/uv/install.sh | sh
source $HOME/.local/bin/env
uv --versioncurl -fsSL https://treg.to/install.sh | sh
treg --versionYou want 0.22.0 or newer, and treg update upgrades an older one. If you see treg: command not found, run uv tool update-shell and open a new terminal window.
treg loginYour first sign-in creates the account. A browser opens with GitHub, Google or an emailed code. Check that the code it shows matches the one in your terminal, and if it asks you to name a team, any name works. If the terminal asks New here? Want a quick setup?, answer n, since the pipeline doesn't need it. No browser on that machine? Run treg login --email you@example.com and type the 6-digit code from the email.
treg org lstreg bills a team, and the pipeline spends from the active one, marked with *. If the list is empty or sign-in said Active org: None, sign in at treg.to in your browser, name your first team there, then run treg login again.
treg balanceNew accounts usually get $1.00 of free credit once, though treg doesn't promise it. If you get it, it covers one first episode of the starter show. After that, the smallest top-up is $10. To add $10, run the command below. It prints a Stripe payment link, and the balance updates seconds after Stripe confirms. Bigger top-ups get bonus credit (5% at $50, 10% at $100, 15% at $200).
treg topup 10One secret does live on your computer: your treg token, in ~/.treg/config.json, readable only by you. Anyone with that file can spend your balance, so never share or commit it. The treg tool also sends anonymous usage stats (command, exit code, duration, OS), which export TREG_TELEMETRY=0 turns off.
Run the installer with the member key from the bot. It downloads the core kit and the starter files, checks them and installs everything into ~/makeroll-ugc-pipeline. Keep the kit out of Desktop, Documents and Dropbox, where a synced file can be moved to the cloud while a run reads it and the run stalls.
curl -fsSL https://makeroll.com/ugc-pipeline/install.sh | sh -s -- YOUR_KEY
cd ~/makeroll-ugc-pipelinePrefer to unpack it yourself? Save the core zip from the bot's message, unzip it into your home folder and run install.sh with your key. install.sh then downloads the starter files by itself.
cd ~
unzip ~/Downloads/makeroll-ugc-pipeline-v2.zip
cd makeroll-ugc-pipeline
MAKEROLL_KEY=YOUR_KEY ./install.shSaved the zip somewhere else? Type unzip with a space, drag the file into the terminal window to paste its path, then press Return. The starter show needs the starter files. Its hand-made character stills and pose plates can't be generated, so without them preflight stops the starter run (exit code 4) before anything is spent.
install.sh checks Python, ffmpeg and ffprobe, a caption font and the treg command line tool from step 2. Run by hand, it asks before it installs anything. The Python packages (numpy, SciPy, Pillow) go into a private .venv folder inside the kit, nothing system-wide, and it never runs sudo without showing you the exact command first. It's safe to run again, and ./install.sh --check only reports.
./ugc doctorIt checks your environment, every file the stages need and the cost model. When something's wrong it names the file or tool and exits with code 6, long before anything can spend money.
./ugc run --dry-runYou get the plan for all eight stages and the estimated spend, worked out from the kit's price table. Nothing executes, nothing is submitted and nothing is charged. It still tells you whether preflight would refuse to start, for example because your treg balance won't cover the estimate.
./ugc runThe stages run in order. If one stops, fix what it names and run the same command again. Finished stages are skipped, and stills or voice lines whose inputs haven't changed come from disk instead of being bought twice.
Exit code 0 means the video passed the 95/100 gate. Exit code 2 means the video exists but scored under the gate, and the report tells you which dimension lost the points.
Preflight says the balance is too low (exit code 4)? Nothing was spent. Check with treg org ls that the team with credit is active, top up with treg topup 10, then run ./ugc run again.
The finished master lands at final/one_prompt_master.mp4 inside the kit folder. The score report, with every dimension and its points, sits next to it at final/quality_report.json.
The kit's starter show is “One Prompt”, a 25-line office skit with two characters, Greg and Nadia. With the starter files installed, their stills, mouth shapes and poses are already built, so the first real run only pays for the voice. The kit's own test run on macOS scored 96.89/100 for $0.0725 of voice, with about 4 minutes of paid stages.
./ugc new my-showThis creates ~/ugc-shows/my-show, a copy of the starter show's inputs (cast stills, mouth shapes, script, voice plan) that shares the kit's engine, so an engine fix reaches every show. Write your lines in production/beats.json, one beat, speaker and line each, then sync the voice plan and check the estimate.
cd ~/ugc-shows/my-show
python3 bin/plan_voice.py --write
./ugc run --dry-runWith the same characters, only the voice lines are new, which makes it the cheapest kind of run. Want your own cast? Run ./ugc new my-show --blank instead. You get an empty template and a checklist of every value still marked TODO in production/project.json, such as the script, the voice plan, the set and style prompts and each character's look.
./ugc statusWhat's built, what's stale, what's broken and who holds the run lock.
./ugc logPast runs with exit code, score, estimate, real spend and wall time.
./ugc costYour balance, spend grouped by endpoint and any paid job still in flight.
./ugc run --from voiceResume from one stage and run everything after it.
./ugc run --only scoreRun exactly one stage. Handy for re-scoring after a manual edit.
./ugc run --forceIgnore the cache and rebuild everything, paid stages included, so it spends credit again. You rarely need it.
MakeRoll doesn't charge per video. Your own treg.to balance pays each paid call at treg's price, and only three stages make paid calls. Cast draws the character and set stills, voice speaks the lines with word timings, and mouths draws each character's mouth shapes. The other five run on your computer for free.
| Stage | What it buys per episode | treg price | Cost |
|---|---|---|---|
cast | 6 images, character and set stills | $0.03 per image | $0.1800 |
voice | 1,642 input characters, dialogue audio + word timings | $0.60 per 10,000 | $0.0985 |
mouths | 42 images, mouth-shape | $0.008 per image | $0.3360 |
| 5 other stages | teardown, script, animate, assemble, score, all on your computer | free | $0.0000 |
| A new cast, first episode | Every paid call, nothing cached yet. A dry-run estimate, not a measured run | $0.6145 estimate | |
| Next episode, same characters | Stills and mouth shapes come from cache, so only the voice lines are bought | $0.0725 measured $0.0985 estimate |
Prices come from treg's catalog and its settled ledger, checked 1 Oct 2026, on the two-character show the kit's starter is built from (the-alignment-series / ep01-im-on-the-call). The next-episode figure is a measured run on 26 Sep 2026, and the kit's starter episode measured the same, 96.89/100 for $0.0725. Preflight wants a balance of at least 1.15 times the high estimate, $0.71 for a new cast's first episode, or the run refuses to start. treg's prices can change, so the dry run is the number to trust on the day.
Before any paid stage starts, preflight compares the estimate with your balance and refuses to begin if it's short. It also refuses any run whose estimate is above your own cap, budget.max_run_usd in project.json ($1.00 for the starter show). An empty balance stops a run at the start, not halfway through a paid job. treg sets its own prices and can change them. ./ugc run --dry-run shows the estimate before you commit, and ./ugc cost shows what you've actually spent.
Every stage declares the files it must leave behind, and the runner checks they're real. An mp4 has to decode with a non-zero duration, a JSON file has to parse and carry its keys, and a PNG can't be a blank fill. Empty output never counts as success.
Loads the reference measurements the final score grades against, and fails early if any are missing.
research/targets.jsonMeasures the screenplay against seven countable comedy defects and refuses a script built wrong, before anything is paid for.
production/beats.jsonDraws the characters and the set. Every still is anchored to one master shot, so faces stay the same from shot to shot.
production/stills/Speaks every beat to its own audio file and saves the exact time each word starts and ends.
audio/beatNN.mp3Draws seven mouth shapes per character in the show's own style, then turns the word timings into a frame-by-frame mouth track.
mouths/Holds each still steady and moves only the mouth and eyelids on top, the way limited 2D animation does.
production/clips/Cuts to the real audio rather than script estimates, burns in the captions and masters the file.
final/Rates the master on eight weighted dimensions. Under 95/100, the run exits 2.
final/quality_report.jsonBefore it scores anything, the scorer fires each detector at a test signal it must catch. If a detector misses its own test, that dimension scores 0 and reads BROKEN. It never reports a pass because it found nothing.
Lip sync carries the most weight. It compares mouth movement with the speech in each shot and is also checked against reversed audio, so a mouth flapping at random can't score.
Every character the pipeline draws and voices is AI-generated and fictional. The starter show's project.json sets a disclosure line, and the score stage checks it's really in the picture. When you publish, switch on the platform's AI-content label too.
Never present a character as a real person, customer, reviewer or testimonial. In the US the FTC's rule on fake reviews bans testimonials from people who don't exist, and EU consumer law has similar rules. MakeRoll's own case pages label every character the same way: an AI-generated fictional character, not a real MakeRoll member, customer or testimonial. The checks MakeRoll runs before anything ships are on the AI video disclosure rules page.
A passing score means the craft checks passed (sync, timing, captions, sound). It says nothing about views, followers or sales. No run promises any of those, or even a passing score.
Licence version 2026-10-01. You can read the full licence before you pay, and it ships in the kit as LICENSE.md. Here's the short version. If the two ever differ, the full text wins.
Install and run the kit on computers you own or control, and keep backups. Change the code, prompts and settings for your own use.
Make videos for your own brand and for clients, as long as you run the kit yourself and deliver the finished videos.
Share, resell, lend or publish the kit or a changed copy, including to teammates, clients or other members. Asking for help with a short excerpt in the private MakeRoll room is fine.
Offer the kit, or anything built on it, as a hosted service, app, bot or API that other people run.
As far as MakeRoll is concerned, what you make belongs to you. The AI providers treg calls for you have their own terms, and whether AI output can be copyrighted depends on where you live.
The licence is personal to you. It ends if you withdraw, if your Lifetime payment is refunded in full, or if it's charged back. Then your member key stops working, and you stop using the kit and delete every copy within 14 days. Your mandatory consumer rights stay as they are.
Anything else goes to support@makeroll.com, answered in writing.
Not natively. treg has no Windows installer and the kit needs bash. The only route is WSL2 with Ubuntu, following the Linux steps inside the Ubuntu terminal, and nobody has tested it yet. For treg login there, open the printed link in your Windows browser or use treg login --email.
No. You type the commands on this page and edit text and JSON files in any editor. Reading an error message helps, and ./ugc doctor names the file or tool that's wrong.
Yes, for your own brand and for clients. As far as MakeRoll is concerned the videos belong to you. You're responsible for your inputs (nobody's likeness, voice or audio without permission), the terms of the AI services treg calls for you, and each platform's AI-disclosure rules. Read the disclosure rules before you publish.
Lifetime Checkout lists two items with their own prices, the kit licence at $149 and the lifetime membership at $50, $199 in total. The kit is final once it's supplied, which is when the bot sends your member key, seconds after payment. At checkout you ask for that and accept that you lose your right of withdrawal for the kit. Within 14 days of payment, /withdraw or makeroll.com/withdraw/ still refunds the unused part of the $50 membership-service share, pro rata. Instead of a refund window you get 90 days of written setup help at support@makeroll.com. The refund policy has the details, including your rights if the kit doesn't work as described.
A full refund, a refund of the kit licence or a chargeback ends your licence and revokes your member key, so the installer, your agent and /pipeline stop working. Stop using the kit and delete every copy within 14 days. A withdrawal from the membership after the kit was supplied doesn't: your licence and key stay.
Only the cast, voice and mouths stages spend, and only when their inputs changed or you pass --force. The dry run shows the estimate first. Preflight refuses to start when your balance won't cover it or a paid job from an earlier run is still outstanding, and paid work is never retried automatically.
The run exits with code 2 and keeps everything. Open final/quality_report.json, find the dimension that lost points, fix that input and run ./ugc run again. Only the stages after your change re-run.
Setup help from Michael Lip for your own brand or site is an optional extra for Lifetime members, on request. It's agreed in writing, no outcome is promised, and you pay your own AI generation costs either way.
Anyone can read it, and every command here is the real one. Only the kit itself is limited to Lifetime members. The bot sends the member key that unlocks it after payment, and agent.md and the installer are public too.
Lifetime is $199 once. It covers the MakeRoll core service for its lifetime, at least 24 months, and a personal licence to this kit, which is separate from the core service. Checkout lists them as two items: the kit licence at $149 and the lifetime membership at $50. AI generation costs are yours, paid straight to treg.to.