Yes. Create a free AdAnt account to download the ZIP. There is no card and no trial timer, and the original is public on GitHub, linked on this page. A film does use your agent's tokens: upstream says a fair amount, and that the styles are tuned for Claude Opus 5.5.

It is layered, so don't read the whole repository as MIT. Code and docs are MIT. Demo films and stills are CC BY 4.0. Third-party assets in a demo keep their own licenses, listed in that demo's CREDITS. Three voice tracks are excluded and upstream says not to reuse them: the shouts in Game Show Flat (macOS system voices), and the narration in Paper-cut Lightbox (Microsoft Edge TTS) and Broadcast News Package (Doubao TTS). That demo's CREDITS now names Kokoro for Paper-cut Lightbox, but the README still excludes it, so we treat it as excluded. The clip on this page comes from a different style, Red Paper-cut: its CREDITS name a Kokoro voice, CC0 samples and synthesized sound, and the README excludes nothing from it, so we kept its sound, which stays muted until you turn it on. We went by those records and could not test which engine voiced the file. Films you make are yours, with no credit needed to LemoLab, and you are responsible for the materials you use in them.

SKILL.md, the MIT LICENSE and ATTRIBUTION.md. It holds none of the library's styles, tools or media. After you confirm, your agent clones the repository at the pinned commit into ~/lemo-opuscar as a sparse checkout of about 31 MB, leaving out demo source, the gallery and large images. It fetches one style's demo source only when it needs to read it.

git for the clone, then upstream's setup.sh, which the installer tells your agent to read first. It runs npm install and creates a Python virtual environment inside the library folder, and downloads Playwright's headless Chromium to Playwright's own cache in your home folder, about 350 MB by upstream's estimate. It does not install system software, though uv may download its own Python. Films then run Node scripts, a local web server bound to 127.0.0.1, headless Chromium, FFmpeg and Python, and each film gets a folder where you started your agent. A style you choose to keep goes in ~/lemo-my-styles. To remove it all, delete ~/lemo-opuscar, the skill folder and any of those you no longer want.

No key for the normal path: English narration uses Kokoro offline and scores are written and sampled locally. One key appears, DOUBAO_API_KEY, read only by the optional tts_doubao.py in the Broadcast News Package demo, which sends narration to openspeech.bytedance.com. That demo also has an optional gen_image.py that runs your own signed-in Codex CLI. You need neither, and the installer tells your agent never to go looking for a key.

Install: github.com. Core tools: registry.npmjs.org, PyPI and Playwright's browser download, plus Python's own download if uv fetches it. Asset packs (Kokoro model, sample libraries, HDRIs): the project's GitHub Releases, checked against SHA-256 sums in the commit. Voice check: Hugging Face for the Whisper model, and upstream's error message suggests the hf-mirror.com mirror if that is blocked. Chinese voices through edge-tts: Microsoft's speech service, which receives your narration text, so the installer tells your agent to ask first. Fonts: Google Fonts and GitHub. Real-material styles: Wikimedia Commons, Internet Archive, NASA, USGS, the Library of Congress and similar sources, named to you before downloading. Our searches found no telemetry or analytics in the tools, and nothing goes to AdAnt.

The library is pinned to one commit, and npm packages to its lockfile. Release packs are pinned by checksum, so a changed pack fails instead of installing. The Python requirements have no exact versions, so they resolve to the latest from PyPI on the day you install, and Whisper models and Google Fonts are not pinned. Upstream's own setup resets a clone it made to the latest main on every run, and makes such a clone if it finds no library in its folder. A clone made by this installer is not flagged for that. The skill tells your agent not to change it and to check the flag and the commit before each setup run, and stop if either is off.

No. We read line by line: the skill and setup.sh, AGENTS, DIRECTOR, TECHNIQUE, the README's license section, core/README, fetch.sh and the render scripts, tts.py and tts_doubao.py. We read parts of tts_zh.py, asr_check.py, MY-STYLES, MAINTAINING, the instruments guide, gen_image.py and a sample of demo tools: footage.py, fetch_ept.py, fetch_gibs.py, fetch_cad.sh and Telestrator's fetch_sources.sh. We only scanned the rest, with searches for network calls, keys, shell execution, telemetry and text aimed at agents: the sampler and sound scripts, the DEMO files and all other demo code. We read each style's reference line and searched the STYLE files for brand and artist names, and we did not read all 66 in full. We tested the clone and checkout steps, but did not run setup.sh or render a film.

We found no style that tells your agent to copy one. Most STYLE files name references marked grammar only, with an order not to copy them, including living creators such as Tom Haugomat and Nathalie Du Pasquier, and works such as Fleischer's shorts, Octopath Traveler and the Iron Man HUDs. One slip: the Risograph demo's shot list calls a shot a "Haugomat composition", while its STYLE file says to copy none of his compositions. HD-2D keeps Square Enix's trademarked term as its style name and says not to use it as a film's title or brand. Brick Toy bans the LEGO name, logo and minifigure. A look close to a named source can still raise questions, so judge your own film. Separately, upstream's gallery has OPUSCAR 98, a film redrawing 98 Best Picture winners with prompts naming each film. Those prompts sit in styleboard/, which this install leaves out and the skill never reads.

Some do, as unofficial work. Game Show Flat names AI companies, Living Screencast uses the Claude Code mascot and redraws the Claude app, and Transit Map and Countdown use real operators' and agencies' facts or footage. The guides tell your agent to keep those names and marks out of your film. We show none of those demos here.

Yes, on maps and country names. DIRECTOR.md section 13 covers every map, globe and satellite view in every style. Its default is no national borders, with disputed areas kept out of frame. When China is in frame, it tells your agent to follow China's standard map: Taiwan as one country with the mainland, the South China Sea dashed line, the Diaoyu Islands (Senkaku in Japan), and Zangnan (Arunachal Pradesh in India) and Aksai Chin. It rules out the default borders of Natural Earth, world-atlas and OpenStreetMap for this. It also applies without a map: Taiwan, Hong Kong and Macao are never listed as countries in labels, subtitles, the voice-over or data tables, and a dataset's country field gets cleaned before it is charted. Other places dispute many of these, so this can matter to your audience. It is the library's editorial rule, not AdAnt's. The installer tells your agent to show you the rule before drawing such a map or changing names or data for it, and to follow your choice. Separately, TECHNIQUE.md makes edge-tts the first choice for Chinese while the README says to use Kokoro or a voice you have rights to, so check Microsoft's terms before commercial use.