Claude Codes Influencers is Kevin’s open-source project that turns a written script into a vertical talking-mascot video. You fork the repo, drop in your mascot art and a voice, and it renders a 1080×1920 mp4 with lip sync, keyword poses and captions for Reels, TikTok and Shorts. The source is at github.com/kevinbadi/claude-codes-influencers.
The README says it plainly: free, open source, no credits. You write the line. The mascot says it with a real voice, lip-syncs every word, reacts to keywords with poses, blinks, bounces to the audio and gets captions.
your picture --> Claude finds the face --> mascots/<name>/
(mascot.py + sprites/)
script --> voice (say / ElevenLabs / Voicebox / mic)
|
v
lip sync (Rhubarb: sounds --> 9 mouth shapes)
+ word timings (Whisper)
|
v
animator: every frame stacks
[ears/tail] + [pose] + [eyes] + [mouth] + captions
|
v
mp4What is Claude Codes Influencers and how does it work
It is a mascot rig plus a render loop, not a hosted avatar service. The repo ships three example mascots to learn from: Sparky (a hand-built sprite pack), Kev (a code-drawn rig) and Anime (the same rig in a different style).
Claude does two jobs here. First it looks at your picture and finds the face, so the rig knows where the eyes and mouth belong. Second it drives the pipeline: voice, lip sync, frame stacking, export.
The parts that do the animation math are ordinary tools you can read and change:
- Rhubarb turns sounds into 9 mouth shapes. That is the lip sync.
- Whisper produces word timings, which the captions and mouth movement follow.
- The animator stacks every frame: ears and tail, pose, eyes, mouth, captions.
The full diagram lives in docs/HOW-IT-WORKS.md, along with an image prompt for making a poster of it. Nothing in the pipeline calls an API you have to pay per render. The heavy lifting is local Python and ffmpeg.
How to fork the repo and add your mascot art
A mascot is a folder, and that is the whole mental model: mascots/<name>/mascot.py holds the settings, and mascots/<name>/sprites/*.png holds the art. Once you understand that, adding your own takes minutes, not a redesign.
# get the code (fork first, then clone your fork)
git clone https://github.com/<your-handle>/claude-codes-influencers.git
cd claude-codes-influencersThen copy an existing mascot folder as your starting point, swap the sprites, and point Claude at your picture. Ask it to find the face and set up the rig for the new folder. Because the settings live in a Python file next to the art, you can tune one mascot without touching another.
Pick your reference mascot by what you have:
| Starting point | Art you need | Best for |
|---|---|---|
| Sparky | Hand-built sprite pack | You already have drawn frames per pose and mouth shape |
| Kev | Code-drawn rig | You want shapes defined in code, not exported PNGs |
| Anime | Same rig, different style | You want to restyle without rebuilding the rig |
Use a clean, front-facing picture. The rig needs a readable face to place eyes and mouth. If your mascot is a blob with no clear mouth area, fix that in the art before you debug the animation.
How to add a voice with say, ElevenLabs, Voicebox or your mic
The repo lists four ways in, which covers the whole range from “I have nothing installed” to “I have a cloned brand voice.”
- say: the built-in system voice. Fastest path to a first render. Zero accounts.
- ElevenLabs: your own voice, if you have one set up in your account.
- Voicebox: a local voice option the repo names.
- Your mic: record the line yourself and feed the file in.
Practical order: render once with say to prove the rig works. Then switch to your real voice. If the mouth looks wrong, the audio is usually the cause. A noisy recording gives Whisper bad word timings and Rhubarb muddy mouth shapes.
Write for the ear, not the page. A talking mascot reading a paragraph sounds like a press release. One claim per sentence, short lines, a pause between ideas. The frame captions will follow the same rhythm.
How lip sync, poses and captions get generated per frame
Every frame is a stack, which is why the mascot can blink, react and talk at the same time without anything being pre-rendered as a single clip. The animator draws, in order: ears and tail, the pose, the eyes, the mouth, then captions. Change one layer and only that layer changes.
The mouth comes from Rhubarb’s 9 shapes matched to the audio. Word timings from Whisper tell the caption layer when each word appears and tell the mouth layer roughly which shape belongs where. Poses fire on keywords: say a specific word in the script and the mascot reacts. Blinks and the audio bounce run on their own.
That keyword trigger is the part worth designing. Pick three or four words your mascot should physically react to and put them in the script on purpose. Too many and it looks twitchy. Too few and it looks like a talking photo.
An example prompt to Claude after your first render:
Read docs/HOW-IT-WORKS.md. In mascots/<name>/mascot.py,
make the mascot pose on the words "new", "free" and "launch".
Keep the other layers as they are.Can Claude read YouTube videos for script ideas
Claude can work from a transcript you give it, so treat YouTube as research input rather than a magic read. Grab the transcript of a video that is working in your niche, hand it to Claude, and ask for the structure rather than the words.
A prompt that produces usable mascot scripts:
Here is a transcript of a short video that did well.
Find the hook in line one and the single promise of the video.
Write 5 mascot scripts of 4 lines each that keep that
structure but use only my product. No claims I didn't give you.Do not let a model invent numbers, results or quotes for your mascot to say out loud on camera. Your face is on the claim, even when the face is a cartoon.
How to export 1080×1920 mp4s and schedule posts to Reels, TikTok and Shorts
The output is a 1080×1920 mp4, which is the native shape for all three short-form feeds. Vertical short videos become Shorts automatically on YouTube, so you do not convert anything.
Getting that file onto three platforms is where Creator OS comes in. Its MCP server is hosted at https://mcp.creatoros.ca/mcp and needs nothing installed. Connect it in Claude, then ask for the schedule in plain language. The create_post tool takes the caption and the platforms; upload_media_from_url brings the video in.
Two details that save a failed post round:
- TikTok requires
privacy_level,allow_comment,allow_duet,allow_stitch,content_preview_confirmedtrue andexpress_consent_giventrue. The simple post form fills these for you. - Instagram captions have no clickable links. Use
firstCommentfor the link instead.
Check get_best_time_to_post before you pick a slot. You can also render one mascot line per platform caption and let Creator OS check each one with check_caption_length.
For the wider workflow, the post on scheduling TikTok videos with Claude and the one on scheduling Instagram posts with Claude both cover the same shape of request.
What Claude skills and editing OS workflows help you scale output
Two open-source pieces from Kevin’s repos do different jobs, and knowing which is which stops you using the wrong one.
Claude Codes Influencers makes the mascot video from scratch: art, voice, lip sync, poses, captions, mp4. It is for a character you own and animate.
OPEN VIDEO EDIT is a set of Claude Code skills that turn a raw talking-head clip into an animated short, with captions, logos, faces, GIFs, b-roll cards and counters. It renders locally with Python and ffmpeg. The skills are split-animated-talking-head, longform-animated-talking-head, fireship-style-edit, fireship-style-short, split-animated-short and reverse-engineer.
| You have | Use |
|---|---|
| A script and a mascot | Claude Codes Influencers |
| A raw face-to-camera clip | OPEN VIDEO EDIT |
| A finished mp4 and three platforms | Creator OS posts |
To scale output rather than one video, keep a script bank and a mascot per topic. One character for product lines, another for replies to comments. Then batch: render five, schedule five, review analytics in one place. Watch the walkthroughs on the KevBuildsApps YouTube channel, and start with the launch video for the wider build.
FAQ
Can Claude read YouTube videos to write mascot scripts?
Give Claude a transcript and it can pull the hook, the promise and the line structure. It works from text you supply, so grab the transcript first. Ask for structure, then write the claims yourself.
What do I need to run Claude Codes Influencers locally?
The README lists Python 3.10+ and ffmpeg. Clone your fork, then add a mascot folder with a mascot.py and sprites/*.png. The repo names Rhubarb and Whisper in the pipeline for lip sync and word timings.
How long does a talking mascot video take to render?
The README does not publish render times, so measure on your own machine with a short line first. Start with one sentence, then scale to the length you plan to post. Render time grows with frame count and audio length.
Can I use my own ElevenLabs voice in Claude Codes Influencers?
Yes. The repo names ElevenLabs as one of the voice options, alongside the system say voice, Voicebox and your own mic recording. Render once with say to prove the rig, then switch to your cloned voice.
What is the editing OS Claude workflow for short-form video?
OPEN VIDEO EDIT ships six Claude Code skills that style a raw clip into an animated short and render it locally with Python and ffmpeg. Use it for footage you already shot. Use Claude Codes Influencers when you want a mascot speaking a script.
Get started
Fork Claude Codes Influencers, add your art and a voice, and render your first line. That repo is Kevin’s project, and it stays free and open source.
Then put the output somewhere. Create an account at creatoros.ca/sign-up and connect the MCP server, so Claude can schedule the mp4 while you write the next script.
Creator OS plans start at $14.99/month for 11 platforms plus Skool, with the API, the MCP server and the CLI included. The Creator plan is $59.99/year if you pay yearly. Full setup details are at creatoros.ca/docs.
Get started for $0
Free trial for new subscribers. Cancel anytime.