~/codigoagente
Open source tools

Slopus: local AI image, video and music generation in one desktop app

Slopus is an open source desktop app that packs an AI agent, local image, video and music generation, reusable references and a chroma-key timeline into a single window. You talk to Claude Code, Codex, OpenRouter or a local OpenAI-compatible endpoint; models run on your GPU through WebGPU, and video is demuxed, decoded, edited and remuxed with WebCodecs and mp4-muxer — no bundled FFmpeg, no Python. It is alpha, ships new builds every few days (this one is 0.4.2, dated November 12, 2026), and the bitti-ai/slopus repository is at commit 34bb69d.

Slopus 0.4.2: agent, local generators, timeline and export in one Tauri app · codigoagente

What Slopus is (and isn't)

  • Repository: github.com/bitti-ai/slopus, published by bitti-ai and marked ALPHA. The current tag is 0.4.2.
  • License: GPL-3.0 for the app itself. You can use it, modify it and redistribute it, but any derivative you ship has to stay GPL and source-available. Model weights carry their own licenses, and several are non-commercial: MiniMax H3 uses the MiniMax H3 COMMUNITY LICENSE, Fibo 2.0 is CC BY-NC 4.0, YuE2 is CC BY-NC 4.0 too, and only MiniMax Music3 is Apache 2.0. If you plan to sell what it generates, read the specific model's license first.
  • Stack: Tauri 2, Rust, React 18 with TypeScript, WebCodecs for audio and video, WebGPU for shaders, mp4box.js to demux and mp4-muxer to mux. The generation runtime is called SlopFab and replaces ComfyUI and Python wrappers.
  • Platforms: Windows x64 and Linux x86_64 (.deb, .rpm and AppImage with in-app updates). There is no macOS build.
  • Real requirements: a GPU in the CUDA family (CUDA 13 for RTX 50, CUDA 12.8 for RTX 30/40) or Vulkan for AMD/Intel. 16 GB of VRAM or more is recommended, and MOSS and Fibo require CUDA no matter what. On Linux you also need WebKitGTK 2.42+ and the NVIDIA driver libraries plus CUDA 13 cuBLAS even when you plan to use Vulkan.
  • Data: projects and media stay on your disk. Agent requests go to the provider you pick; API keys are never saved inside the project.

What it is not: not a cloud assistant, not a Midjourney clone and not a drop-in replacement for Premiere. It is an authoring harness to try local generation flows with an agent beside you.

What's in the box

Slopus merges four apps that would usually sit in separate windows:

  • Images. Batch up to 100 images with MiniMax H3, MOSS Image, Krea 2 Turbo/Raw, or Fibo 2.0 Base/Edit Turbo. Supported models accept whole-image, box, lasso and Zoom edits, plus character sheets. Bounded edits keep the original resolution and store drafts and versions.
  • Video. Projects live as Videos → Scenes → Shots. Each scene has a type: Prompt, Single Shot, First & Last Frame, Backdrop (green/blue/black/white background keyed out later), Continue (continues from another scene's saved latents), Animate, Pose, Character Replace, Extend and Bridge (connects two videos). Each shot keeps four takes plus a draft, you pick the main take and can upscale the saved latents to 1.5×, 2× or 4×.
  • Music. Generate with YuE2 or MiniMax Music3, edit lyrics and style, organize songs into folders and play them in sequence. Right-click a song to generate its album art. Song variations (0.4.1) duplicate a song as a draft and queue up to 100 variations with different seeds.
  • Timeline. Drag video and audio clips, trim, split, snap, read waveforms, apply transitions, chroma key and GPU effects, and plug in Motions — JavaScript animations the agent can write, edit and render as clips or references.

All of this shares a top bar with Save, Undo and Redo, project settings (slopus.json inside the folder) and an Export pane that writes MP4, JPG/PNG, or audio as WAV, MP3, lossless FLAC, AAC/M4A or Opus (the last two depend on OS support). MP3 and FLAC encoders ship inside.

The agent is the thing that sets it apart

The agent pane opens with Ctrl+Shift+A. Per docs/architecture.md, each provider is a subprocess adapted to a shared request/event/result contract. Claude Code runs non-interactively with claude --print --output-format stream-json (never --bare, which ignores OAuth and keychain logins and rejects any subscription session). Codex runs with codex exec --ephemeral --ignore-user-config --json. OpenRouter or a local OpenAI-compatible endpoint live under Settings → Agents.

The security pinch matters: the agent receives the full project as read-only context, can reply, ask a question or propose a JSONL command stream ending in commit. Slopus parses it into a typed Rust enum, dry-runs the batch, hands it back to the editor, and the editor re-runs it against the latest state before saving once. A failing command applies nothing, and providers never get authority to write arbitrary files. That's why you can let it work without fearing it will nuke scene 7 because it misread the prompt.

A minimal flow with Claude Code

  1. Install the right release from GitHub Releases. On Windows pick the x64 installer; on Linux pick .deb, .rpm or the AppImage (chmod +x).
  2. In Settings → Agents pick Claude Code (if claude is on your PATH) or configure OpenRouter with your key.
  3. In Settings → Generators, pick a model and hit Download weights until the generator reads Ready. Weights are big: 10–30 GB per video model.
  4. Create a project, open the agent pane with Ctrl+Shift+A and ask something like «plan three ten-second scenes for a manual coffee-maker ad, with a Bridge between scenes 2 and 3».
  5. The agent writes the storyboard, proposes the commit, Slopus validates it, and your scene list is ready for Generate All.

If the editing machine has no GPU, jump to LAN workers.

Included models and what they do

ModelKindNotes
**MiniMax H3**Video + imageCore of the pipeline. Its takes feed every scene type. `Character Replace` needs reference-capable weights (`References` or `Singularity`).
**MOSS Image**ImageRequires CUDA. Sharp for illustration.
**Krea 2 Turbo / Raw**ImageFast editing with box and lasso.
**Fibo 2.0 Base / Edit Turbo**Image**Non-commercial** (CC BY-NC 4.0). Requires CUDA.
**YuE2**Music**Non-commercial** (CC BY-NC 4.0). New default: 6 flow steps with DPM++ 2M.
**MiniMax Music3**MusicApache 2.0. The only out-of-the-box commercial-friendly model.

Settings → Generators shares weights, LoRAs, attention, generation effects and compatible sampler/scheduler pairs (ER-SDE and Beta included). It warns you when VRAM is short and shows live previews for H3 images and video.

LAN workers (when your laptop has no GPU)

This is the piece that turns Slopus into something useful for most people. The editor runs wherever you want; weights and generation live on a LAN PC with a GPU.

LAN worker: laptop without GPU, GPU desktop serving generation · codigoagente

On the GPU PC, unpack Slopus-<version>-windows-x64-worker.zip or Slopus-<version>-linux-x64-worker.tar.gz and launch it:

terminalsh
./slopus-worker --weights ~/models

Or on Windows:

terminalsh
slopus-worker.exe --weights D:\Models

Open TCP 47321 and UDP 5353 (mDNS) in the firewall. In Slopus, head to Settings → Workers and click Use next to the discovered worker; or add its address by hand. Missing weights download themselves if the link is reachable. To go back to local, pick This computer.

slopus-worker --help lists more options: backend, access token and port. Linux workers need the same runtime libraries as the editor.

What it costs

ItemPriceNotes
Slopus**0 USD**GPL-3.0, free installer and AppImage.
Model weights**0 USD**Downloaded from `Settings → Generators`. Watch the license: Fibo 2.0 and YuE2 are **non-commercial**.
Recommended GPU400–1,800 USDAn RTX 4070 Super (12 GB) will struggle; an RTX 4090 (24 GB) is more than enough. Under 16 GB of VRAM locks you out of H3 and Fibo 2.0 Base.
Agent (optional)0 USD on Claude Code with your subscriptionOpenRouter is pay-per-token; `Codex` only needs a free login.
Rented GPU0.20–0.80 USD/hourRunPod or Vast.ai on an RTX 4090. Fine for test rounds, not for steady production.

For context: Midjourney's Basic plan is 10 USD/month, and a separate commercial license for Premiere + Audition is 22.19 USD/month. In Slopus the only real recurring cost is your GPU's electricity; the trade-offs are the up-front hardware price and the model licensing ceiling if you want to sell the output.

If you have no GPU and don't want one, two options: rent one by the hour (great for trials, rough for continuous production) or keep only the agent + OpenAI-compatible endpoint path and use Slopus for the timeline, references and motions, bringing in generations produced elsewhere.

The good, the raw and the ugly

The good. The agent is more than a chat: it proposes a batch of commands validated against the project schema. The WebCodecs + WebGPU timeline is quick and does not force a bundled FFmpeg with its GPL/LGPL baggage. The scene contract (Continue, Bridge, Pose, Character Replace) covers real production cases. And the .slop bundle format for generator templates plus the LAN workers give a small studio a clear path to scale.

The raw. It is alpha and it shows. There are constant fixes (0.4.2 patches MiniMax H3 resolution, character sheets and Zoom), file formats still read from previous versions (polstudio.json, pols.json), and corners that warn about incompatibility when you change a scene type. If you fold it into a workflow, expect updates every few days and keep backups.

The ugly. No macOS. Fibo 2.0 and YuE2 licenses make the common agency case («sell this output») impossible with those models. And the 16 GB VRAM recommendation locks out ordinary laptops: without a strong GPU or a rented one, the «local generation» pitch loses half its charm.

When to pick Slopus

Pick Slopus if:

  • You have a strong GPU (16 GB of VRAM or more, ideally) or access to a LAN PC with one.
  • You want to experiment with references, motions and conditioned scenes without hopping between ComfyUI, DaVinci Resolve and an audio editor.
  • You like a flow where Claude Code or Codex propose scenes and you approve by batches.
  • The content is personal, internal or creative without sale, or you only plan to run models with a compatible license (MiniMax Music3 and whatever you bring in).

Skip it if:

  • You work on macOS and don't want to run a VM.
  • Your business sells the output and you're not paying for or switching to commercial models.
  • You have no GPU and don't want to rent one: the frustration curve with cloud is real.

Keep going

If you try it, note which weights worked for you and which version you were on: Slopus moves fast and release notes thank the feedback.

// source code

Repository in this article
The code this article covers is public: browse it, star it or clone it from here.

Help us grow

codigoagente is free, with no paywalls. If this article helped you, any of these keeps the blog going.

// newsletter

One email a week with the repos worth your time

The week's new articles, tested against real code. No spam; unsubscribe in one click.

ArtCraft is an open source app for AI images and video where you build the 2D or 3D scene first. Features, real costs and how to use it with your fal key.
  • Free alternatives to paid apps
  • Design, image and video
read article
PeakCut cuts long videos and YouTube, Twitch or Kick streams into 9:16 clips with word-by-word captions. Install guide, real costs and a Spanish-language test.
  • Free alternatives to paid apps
  • Run it on your server
read article
Breakscale is an MIT system design simulator: build your architecture, raise the traffic and watch why it fails. Free on the web, in VS Code and via MCP.
  • Claude Code
  • Free alternatives to paid apps
read article