Locally Uncensored: a Tauri desktop studio that wires Ollama and ComfyUI into one window
The all-in-one local AI studio for your desktop: chat, image and video generation and a coding agent in one free, open source app. Windows and Linux. No Docker, no terminal, no cloud required.
At a glance
- What is it?
- Locally Uncensored bundles chat, image and video generation, a coding agent and a tool-using agent into a single Windows and Linux desktop app, with AGPL-3.0 licensing and an optional hosted GPU tier. The value is the packaging, not the models, and the packaging is also where the constraints live.
- Who is it for?
- Adopt it if you want local chat, ComfyUI image and video lanes, and a diff-first coding agent behind one installer on Windows or Linux, and you accept AGPL-3.0 terms. Do not adopt it if you need macOS support today, if you cannot run a local engine plus ComfyUI, or if your workflow depends on the two cloud-only lanes.
- Can I use it commercially?
- Yes, with strict conditions. AGPL-3.0 is a network copyleft licence: if people use a modified version over a network, for example as a hosted service, you must offer them its source code under the same licence.
- Is it still maintained?
- Yes. The repository last received commits 1 day ago.
- What is it written in?
- Mainly TypeScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The problem is installation, not inference
Running a local model is no longer the hard part. Wiring a chat UI, a ComfyUI instance, a model downloader that respects your VRAM, and a coding agent into something a non-specialist can launch is the hard part. Locally Uncensored targets exactly that gap. The README frames the product as an all-in-one studio where you install it, pick a model and go, with no Docker, no terminal and no config files. The audience is the person who already has a GPU and does not want to hand-assemble four separate projects to use it. The repo is TypeScript on a Tauri shell, Windows and Linux only, and the README states macOS has no build yet and must be compiled from source with npm run tauri build.
One window, four tabs, two external engines
The architecture described in the README is a desktop shell that orchestrates engines it does not ship. On first launch a wizard looks for an AI engine already running on the machine and installs one if none is found, and does the same for ComfyUI, which the app installs, starts, repairs and updates itself. Chat, Create, Code and Agent are tabs in one window sharing the same model selection. The Create tab exposes lanes rather than node graphs: text to image, edit, image to image with a mask for inpainting, remove background, video, animate an image, extend a clip, motion control, talking character, music, and Character Studio, which trains a character and writes it into your local LoRA folder. The Code tab builds a map of the repository, shows a diff before applying edits, runs tests and reads failures, and reads per-project rules from a .lurules file. The Agent tab adds web search, file read and write, a shell, code execution, screenshots, and user-supplied MCP servers, with a permission gate on every tool call and a read-only mode.
Getting it running without touching a terminal
The normal path is the installer: an .exe (NSIS, which the README recommends) or .msi on Windows 10 and 11, and .deb, .rpm or .AppImage on Linux. The first-launch wizard handles engine detection and ComfyUI setup. From source, the README gives four commands: git clone the repository, cd into it, npm install, then npm run dev for browser dev mode or npm run tauri build for the desktop binary. On Windows setup.bat bootstraps Node, Git and Ollama for dev mode, and setup.sh does the same on Linux and macOS. There is one config surface worth knowing: a Local API toggle in Settings that exposes every local model on one OpenAI-compatible address behind a token, so other coding tools on the same machine can point at it. Project rules live in .lurules at the repository root.
Two Create lanes leave your machine
The README is explicit that Upscale and Erase Object are the two lanes that run in the cloud only. That is a real boundary for a project whose pitch is that everything runs on your own machine, and it deserves to be read as a limitation rather than a footnote. There is also LU Labs Cloud, an optional switch for models larger than any desktop card, with heavy Create lanes alongside them, an account that also works at lu-labs.ai on a plan or on credit packs. The README states the local app stays free either way. The practical consequence is that a fully offline workflow is possible, but not across every lane the UI presents, and the app will not tell you that at the point of use unless the switch state is visible.
Unsigned installers and what you can actually verify
The README addresses antivirus flags directly: some engines flag unsigned NSIS installers that download other binaries, which the project calls a false positive. That claim is checkable rather than takeable on faith. The installer is built by GitHub Actions from public source on master, and the update channel is signed against a public minisign key, so SECURITY.md is the document to read before you run anything. If you are deploying to a managed fleet, the combination of an installer that fetches other binaries and a desktop app that can start a shell and execute code through the Agent tab is a policy question, not just a convenience question. The permission gate and read-only mode are the controls the README names; how they map to your environment is something you have to verify yourself.
Where a hand-built stack still wins
The obvious alternative is the assembly the app replaces: Ollama or LM Studio for chat, ComfyUI directly for image and video, and a terminal coding agent such as Aider or a similar tool for repository edits. That route costs setup time and gives you the node graph, the raw ComfyUI API, and full control over which custom nodes and workflows exist. Locally Uncensored deliberately hides that layer, which means a workflow that depends on a specific custom node or a hand-tuned graph may not be expressible through the lanes the README lists. Conversely, if your only goal is chat against a local model, the app is heavier than the engine alone, since it also installs and manages ComfyUI. The trade is setup effort against control, and the README does not claim the lanes cover every ComfyUI capability.
Release cadence, licence and the cost of staying current
The release history shows a fast cadence: v2.6.7 on 31 August 2026, v2.6.8 on 6 September, v2.6.9 on 8 September, all within roughly a week, with the changelog tracking every change since 1.0.0. Frequent releases plus a self-updating signed channel means upgrades are low-effort but also that behaviour can shift between minor versions, and that the app is managing ComfyUI and engine versions underneath you. Budget for that as ongoing maintenance rather than a one-time install. The licence is AGPL-3.0, which carries obligations if you modify the code and make it available over a network; if you plan to embed or redistribute it, read the licence text and take your own advice rather than treating this as legal guidance. The README also notes the local app remains free regardless of the cloud tier.
Editorial conclusion
Adopt it if you want local chat, ComfyUI image and video lanes, and a diff-first coding agent behind one installer on Windows or Linux, and you accept AGPL-3.0 terms. Do not adopt it if you need macOS support today, if you cannot run a local engine plus ComfyUI, or if your workflow depends on the two cloud-only lanes. Before committing, verify the minisign key against SECURITY.md, confirm your GPU can host the ComfyUI models you intend to use, and check whether your existing Ollama or LM Studio model store can be linked rather than copied.
Community notes