01What Is the Open Source Higgsfield Called Treg?

The open source higgsfield is a tool called Treg that lets your AI agent make AI videos for your brand from inside Claude Code or Codex. You type one message, and your agent finds a winning video format, builds an AI creator, picks a voice and makes the finished video.

Treg launched on X on the 13th of November. The launch post called it Open Source Higgsfield and said three things: it copies any viral video format, it runs directly inside Claude Code or Codex, and it is open source. The demo showed a giant cat walking down a city street while two creators filmed a selfie video in front of it.

At its core, Treg is a tool catalog for your AI agent. The team describes it as being like OpenRouter, but for tools instead of models. OpenRouter is one connection that reaches lots of AI models.

Treg does the same for tools. You do not sign up to each provider or juggle 20 logins. Your agent asks for the job it needs done, and Treg finds the tool that does it.

Because the code is public on GitHub, anyone can read it, check it, or run their own copy.

02Main Features of Treg and the AI UGC Video Skill

Julian Goldie explains in this episode that the part of Treg getting all the attention is AI video. Treg ships a skill called AI UGC video. UGC means user generated content.

It is the style of video that looks like a normal person filmed it on their phone.

That style works well in ads right now because it does not look like an ad. Think of a creator holding your product, talking straight to the camera, or walking down the street. The skill is a playbook your agent follows step by step, so it makes these videos properly instead of guessing.

Here is what Treg gives you:

  • One connection to over 4,000 tools from 127 providers.
  • Video models, image models, voice models, social trend data and research tools.
  • A step-by-step video playbook that runs inside your agent.
  • A library of over a thousand English voices from Fish Audio.
  • Your own provider key used first, if you already have an account with a provider.
  • Public code you can read and run yourself.

For an online store, that means product videos that look like a real person filmed them. For an agency, it means a batch of ad videos for a client in three versions you can test.

03The Four Steps the Open Source Higgsfield Runs

The skill runs four steps, and your agent does every one of them from inside Claude Code.

  1. Research. The agent searches social platforms for the most liked posts in your niche and picks one as the reference. It then breaks that video down: what is on screen in the first three seconds, each beat second by second, how the camera moves, where the light comes from, and who talks and when.
  2. Your AI creator. The agent makes brand new characters for the roles in the reference video. They are not the real people. For each one it builds a character sheet, which is one image showing the person from the front, side and back, plus close-ups of the face. Then it makes the empty location and any props, like your product.
  3. The voice. The agent browses the Fish Audio library. You hear the line and pick the voice. The voice is handed to the video model as a reference so the lip sync follows it.
  4. The video. Everything goes to Seedance 2.5 together: the reference video, character sheets, location, product and voices. Out comes your video in the same format as the viral original.

The research step is the smart part. A post that already went viral shows you the framing, pace and kind of people that work in your niche, so you start from something already proven.

04How the Demo Video Was Made, Step by Step

The launch demo shows the whole flow on one example. The agent found trending content, filtered creators, built an AI creator from scratch and picked voices from the library. For the giant cat video, the team picked a voice called Chloe for one creator and a different voice for her friend.

The voice gets its own step for a reason. It sells the character as much as the face does. A good face with the wrong voice still feels off.

Then the team did something very useful. They made three versions of the same video:

  • One in a purple setting.
  • One on a city street.
  • One at golden hour on a rooftop.

Same format, three looks. That means you can run all three as ads and see which one people stop for. The demo also ended with a side-by-side of the original video and the new version, shot for shot.

Julian points out that within hours tens of thousands of people had watched the launch post. The interest came from seeing the full job done in one conversation: research, characters, voice and the final video, with no website to click through.

05Open Source Higgsfield vs Higgsfield: Five Differences

Higgsfield is the popular AI video app. You log into the website, pick a preset, upload an image and click generate. It has camera moves, character tools and access to many top video models.

Julian is fair to both, and compares them in five ways.

  1. Where you work. Higgsfield is a website you click around in. Treg lives inside your AI agent, so you never leave Claude Code or Codex.
  2. How much of the job it does. Higgsfield makes the video when you ask. The research, script, characters and plan are mostly up to you. The Treg skill runs the whole job, start to finish.
  3. What else it can reach. Higgsfield is a video and image studio. Treg reaches over 4,000 tools, so the same agent can look up trends, research competitors and pull social data.
  4. The models. Both reach many of the same video models, and Seedance 2.5 is on both. Treg also lists Wan and Minimax for video, plus Nano Banana Pro, GPT Image and Flux for images.
  5. Open source. You can read exactly what Treg does and run your own copy. Higgsfield is a closed app.

Higgsfield still wins if you never want to open Claude Code. It has a polished interface, lots of presets, and you can see everything visually.

06How to Set Up Treg in Claude Code or Codex

Setup takes a few messages, and you do not write any code. Julian walks through it like this:

  1. Open Claude Code or Codex.
  2. Paste the first setup message into your agent. You can sign in with GitHub, Google, or a code sent to your email.
  3. Paste the second message, which tells your agent to follow the Treg AI UGC video skill file to make you an AI video from a viral reference.
  4. Answer the agent's questions. It asks about your niche, your product, your characters and your voices, one question at a time.
  5. Review the plan. Before it makes any video, the agent shows you the plan and asks you to approve it.

That last step is the checkpoint. Nothing gets made until you say yes.

One tip matters more than any other. Settle everything at the image stage first. If the face or the product looks wrong, fix it there before you ask for any video.

Julian calls this the single biggest thing that makes the final video look right the first time.

07Built-In Rules That Make AI UGC Video Look Real

The skill has rules built in, and they are good ones. They are the reason the videos look like phone footage instead of a glossy render.

  • Camera details first. The agent writes the camera details first in every prompt, because that is what makes AI video look real.
  • Real flaws on purpose. It asks for a little handheld wobble, focus that hunts for a second, and visible skin texture.
  • Draft first. It makes a quick, low resolution draft and only makes the full quality version once you are happy with the draft.
  • Honest about audio. It tells you it cannot hear audio, so you check the voices yourself before the final version.

A few mistakes to avoid:

  • Do not skip the image check. A wrong face or product gets harder to fix once it is in video.
  • Do not skip listening to the voices. The agent cannot do that part for you.
  • Do not test everything at once. Julian suggests one AI creator, one location and one video, tested against your current best ad.

08Who Treg Is For: Stores, Agencies and Creators

Julian compares the old way and the new way. The old way means scrolling for hours, messaging creators, shipping them your product, waiting for changes, and ending up with one version. The new way means telling your agent your niche and product, and getting three versions to test in one afternoon.

Here is how different businesses can use it:

  • Online stores. Say you sell a skincare serum. Ask for an AI creator holding it in a bright bathroom with window light. Selling hoodies? Ask for two friends walking toward the camera, in three different places.
  • Product launches. Your agent can make the launch video the day you have product photos, so ads are ready before the stock lands.
  • Brand faces. Because the creator is locked, the same AI creator can appear in every ad, so customers start to recognize the face.
  • Agencies. Offer it as a new service, make case study videos with an AI presenter, or use the research step before a client call.
  • Content creators. Take one idea and turn it into a short video in a format already winning in your niche.

If you sell anything you can show on camera, or run an agency that makes ads, Julian says this fits.