troupe

A video studio that runs in your browser.

Write a short script, cast one of 30 actors, ask the chat for a sharper hook, then render an MP4 with a real voice. Open source, with no account, and your projects stay on your device.

In your browser

  • The whole studio, its database and your renders live in this tab.
  • Kokoro voices and captions, rendered in the tab: on your GPU when the browser has WebGPU, on the CPU otherwise.
  • A script chat that runs on your GPU too: no key, no server.
Open Troupe in your browser

Chrome or Edge. The voices download once, on your first render: 326 MB with WebGPU, 92 MB on the CPU.

One project, from an empty studio to the downloaded MP4, recorded in Chrome. Waits are sped up, and the video says so.

On your server

  • Cloud video models with your own keys: Veo, Kling, Seedance.
  • Your own GPU through ComfyUI, or the local renderer.
  • Keys stay on your server, encrypted at rest.
Self-host with Docker

One docker compose up. Cloud providers bill your own account.

From a script to an MP4

The project in the video, step by step. Each step is a page of the studio, the same in both editions; the times jump the video to it.

  1. Start a project

    Pick a platform, a format and a language. The model picker only offers what each model can do, and says why before you launch. Then cast an actor.

  2. Write the script

    A line per sentence, each with an emotion that sets the delivery. Every save is a version you can restore.

    1. Mornings are hard.hook · neutral
    2. Our cold brew is smooth, strong and ready in your fridge.body · neutral
    3. Grab a bottle on your way out.cta · neutral

  3. Ask the chat

    Ask for a change in your own words. The answer is a whole new version, shown line by line against the current one, and nothing changes until you apply it. In the browser a small model answers on your GPU; a self-hosted studio uses Ollama or Claude.

    Make the first line punchier. Keep the other lines.

    1. Mornings are hard. But not for us!

      Mornings are hard.

      hook · neutral excited

    2. Our cold brew is smooth, strong and ready in your fridge.

      body · neutral

    3. Grab a bottle on your way out.

      cta · neutral

    Longer than the 9 s clip. Ask for a shorter version, or relaunch on a longer clip.

    The proposal as the video shows it: a new hook, delivered excited, and the other lines kept as asked. It runs past the 9 s clip, so Apply & relaunch renders it on a longer one, and the new take lasts 11.3 s. You read every proposal before you apply it.

  4. Render it

    In the browser edition, Kokoro voices each line and the frames are drawn and encoded in the tab. A self-hosted studio sends the same project to a cloud model, ComfyUI or the local renderer, and follows its progress.

  5. Watch it, then download it

    The render plays on the project page with its voice. Every render is a draft until you export it; the MP4 is named after the project, the model and the time.

    cold-brew-launch-kokoro-voice-captions-2026-10-06-2006.mp4
    720×1280 · H.264 + AAC · 11.3 s

Thirty actors, six pictures each

Synthetic people, generated for Troupe with FLUX.2 [klein]: nobody was photographed. Each has a voice and six pictures: a front view, two profiles and three expressions. A render’s actor card shows the front picture, or the happy, calm or excited one on lines with that emotion; the AI video mode draws no card. With a cloud video model the actor guides the prompt, and the face changes from one render to the next.

Two editions, the same studio

The same pages and the same router. What one edition cannot do is marked where it would be, with the reason.

In your browser On your server
Where it runs This tab: Postgres compiled to WebAssembly, kept in the browser’s storage. Docker on your machine or a server, with PostgreSQL. Vercel and Supabase work too.
Your data On this device only. Export a backup to move it; clearing the site’s data deletes it. Docker volumes on the server: the database, the videos and the key that encrypts saved API keys.
Video models Kokoro voice + captions, rendered in the tab. Veo, Kling and Seedance with your keys; LTX-2 and Wan through ComfyUI; the local renderer, with an opt-in AI video mode; any HTTP model server.
Cloud models Not here: an API key needs a server to keep it. Billed by the provider to your own account.
Script chat Qwen2.5 1.5B through WebLLM, on your GPU (880 MB, once). Ollama, running in the Docker stack, or Claude with an Anthropic key.
Model servers on your machine Not here: a page from the web cannot reach them. ComfyUI, the local renderer, Ollama, or your own.
While it renders Keep the tab open: the render runs in it. Close the page; the server checks on the job.
To start Open Troupe in your browser Self-host with Docker

Self-host with Docker

One command on a machine with Docker, nothing to configure. The first start downloads the chat model and the voices (about 2.6 GB), then the studio opens with the local renderer and the chat ready. Its access code is in docker compose logs app.

Self-hosting guide · Vercel + Supabase

Quick start 
git clone https://github.com/maxgfr/troupe.git && cd troupe
docker compose up -d --wait

Models, and where they run

Each model declares its formats, lengths and audio; the studio only offers those, and says what it has and has not been tried on.

Limits, plainly

Open source

Troupe is MIT licensed, actors’ pictures included. Some parts it uses carry other terms:

eSpeak NG, in the browser edition
The render worker bundles phonemizer, which embeds eSpeak NG compiled to WebAssembly. eSpeak NG is GPL-3.0-or-later, so the built site combines MIT code with GPL code; its source is on GitHub.
Model weights
Kokoro-82M and Qwen2.5-1.5B-Instruct are Apache 2.0. Your browser downloads them from Hugging Face; Troupe does not distribute them.
LTX-Video, in the opt-in AI video mode
The weights use the LTXV Open Weights License, which is not an open-source license: free under $10M of annual revenue, and bound to use restrictions, among them saying that content is machine generated and not impersonating anyone.
FFmpeg, in the Docker image
GPL, run as a separate program.

Every third-party notice