TensaikunYT/Videohub

What it does

VideoHub is a local desktop studio for videos that an AI coding agent writes as code. A film is a web page that draws every frame in JavaScript via window.renderOut(t), so VideoHub can play it live without rendering, step through frames, and let you draw annotations and leave notes on any frame or frame range. Notes are exposed to Claude Code or OpenCode through MCP tools (hub_notes, hub_note_update, hub_live, hub_render, and others), so the agent can read them, edit the code, and let the preview reload in place. Rendering is done by an ffmpeg-based queue that processes one job at a time, with partial re-renders of only the 2-second chunks your notes touched.

The tool targets people who generate video through a coding agent: motion graphics, narrated explainers, and similar frame-by-frame work. It ships with agent tooling, a sound mixer, a thumbnail studio, and render-protection controls, and can run the same hub_* tools from a shell via node server/cli.js.

Getting started

Prerequisites on every platform are Node.js 22+, ffmpeg on PATH, a native Chrome/Chromium/Edge build for headless rendering, and Claude Code or OpenCode. Windows 10/11 can install the first two with winget; Linux and macOS use the distro/Homebrew packages described in the README requirements. Snap or Flatpak browser wrappers are not supported.

1. Download the ZIP or clone the repository into a folder you want to keep.
2. Windows: double-click SETUP.cmd (preview with setup.ps1 -DryRun). Linux/macOS: run chmod +x setup.sh update.sh uninstall.sh open.sh, then ./setup.sh (preview with --dry-run). Setup asks about workspace, watched folders, GPU encoding, agents, and the example film.
3. On Windows launch from the Desktop shortcut or open.cmd; on Linux/macOS run ./open.sh. Restart the agent session so it picks up the hub_* MCP tools.
4. Open the example film, press Space to play and ← → to step, click a frame to leave a note, then Render → Full film. Expect hello-film_v1.mp4 in the workspace; after an edit, Render → Only the fixed parts writes hello-film_v2.mp4 and can play only the changed ranges.

The new-film template and project contract describe what a film folder must contain for live preview, notes, and chunked partial renders.

Constraints and security

A graphical desktop session is required for windows, folder dialogs, notifications, and file-manager actions. Linux needs a native browser and optional taskset, zenity/kdialog, and notify-send; without taskset or on macOS the CPU-affinity power controls are unavailable. The render queue executes shell commands for your projects, so the server binds to 127.0.0.1 only and should never be exposed on port 7788 via tunnels or port forwarding. Setup and update scripts lift the PowerShell execution policy only for their own process; if a machine policy still blocks them, use an approved signed deployment instead of lowering protections.

Linux and macOS support is described as new and not yet tested on real machines, and Linux/macOS render performance is unverified; the platform notes and render speed document feature limits and existing Windows measurements. The project is MIT-licensed.

HotRepo's take

The project entered HotRepo's new-project list on 2026-11-13 with 108 stars, 29 forks, and 6 stars gained over a one-hour window. That is a short observation, so it does not establish a sustained trend. The notable engineering decision is the code-as-video contract: preview, notes, and partial re-renders all depend on an explicit page interface, which makes the pipeline workable without compiled native helpers. The practical tradeoff is setup weight—Node, ffmpeg, a native browser, and an agent—and the still-unverified Linux/macOS rendering path.