Model or dataset
PurpleDoubleD/locally-uncensored avatar
PurpleDoubleD/locally-uncensored

locally-uncensored: Windows is tested on every release, Linux is only built

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.

4,539 stars728 forksTypeScriptAGPL-3.0

At a glance

What is it?
A desktop studio that runs chat, image generation, video and a coding agent on your own machine, wiring an existing local engine and ComfyUI together so there is no Docker and no terminal. The release notes are the most useful part of the file, because they name the failures: long file writes ending runs, chats running the app out of memory, and an AppImage that used to hand its bundled runtime to every program it launched.
Who is it for?
This suits someone who already runs a local engine and wants one window instead of four tools, and the linking of existing model directories rather than copying them is the detail that decides whether the disk survives. Three things to know.
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 November 21, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • PurpleDoubleD/locally-uncensored4,539 ★ · +2,989
Star history
Repository2026-09-102026-11-21
PurpleDoubleD/locally-uncensored1.6k4.5k

2,989 Stars in the last 72 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

Windows is tested, Linux is built, macOS is absent

The download table carries a status column that does the real work. For Windows 10 and 11 there is an NSIS installer, recommended, or an MSI, and the status reads tested every release with a signed auto update channel. For Linux there are three artefacts, an AppImage, a Debian package and an RPM, and the status is more modest: built on every release. For macOS the row says no desktop build, and the substitute is the cloud edition running in a browser at a separate address. So the difference between the two desktop platforms is not cosmetic. One is exercised before it ships and the other is compiled, which means a Linux user is closer to being the first person to run the build they downloaded. The file names exact filenames for each artefact, so an installer can be traced to a specific version without opening a release page.

Antivirus warnings answered with a minisign key

One paragraph in the download section deals with a problem almost every self-contained desktop installer has: some antivirus engines flag unsigned NSIS installers that download other binaries, which the file calls a false positive. The answer it offers is verifiability rather than a code signing certificate. The installer is built by GitHub Actions from the public source on the master branch, so the artefact corresponds to code anyone can read, and the update channel is signed against a public minisign key, so the updater can be checked independently of the installer. Both checks are pointed at a security document in the repository with a section named for antivirus and browser false positives. That is a reasonable answer for a project that cannot afford commercial signing, and it is also an answer that depends on the reader actually verifying: an unsigned installer from an unsigned channel is otherwise indistinguishable from the thing antivirus warns about.

Four patch releases, two of them admitted as hotfixes

The release notes are the most substantive documentation in the file, and they read as a bug list rather than a feature list. Version 3.0.2 was a hotfix for the code tab losing its scrollbar and input box once a conversation grew taller than the window, described as a regression in 3.0.1, and it also widened the cloud model count to fourteen video and seven image models by including the create studio endpoints. Version 3.0.3 was a hotfix for chats with many images running the app out of memory, and for the create flow reading a model's family from the file rather than from its name. Version 3.0.4 is the current release and concentrates on the agent: writing a long file no longer ends the run, continue resumes where a run stopped instead of starting over, the agent keeps a plan only when asked for one, and cloud context defaults to 32K and is settable per model. It also lists five community-reported bugs by issue number.

The AppImage stopped shadowing your system Python

One entry from the previous release explains more about how the app is packaged than any feature list would. The Linux AppImage used to hand its own bundled runtime to every program it started, which meant a healthy system Python, pip, git or ffmpeg could stop working while the application was running. The fix stops that inheritance, with an exception in the other direction: a Linux distribution that blocks pip inside the system Python keeps the isolated environment the app built, so the fix is conditional on the host being sane. Two more entries in the same release are quietly consequential. A GPU whose free memory cannot be measured is now planned more carefully, and the log says total capacity instead of calling it free, which is a small correction with real consequences for a model manager. And a long conversation folds its older turns into a summary, which changes what the model sees in a way users will notice and cannot inspect.

One local conversation at a time, with the queue shown

Concurrency is not uniform in this application, and the file says which part is limited. The local model runs one conversation at a time and tells you so, with a line showing how many chats are ahead of you, and stopping works while a chat is still waiting. The cloud side has no such limit, which is the trade the product is making: local means private and serial, cloud means parallel and remote. Settings follow the same pattern. Temperature, top probability and maximum tokens belong to the conversation you moved them into, rather than every open chat sharing one value, which fixes a class of confusing bug where changing a setting in one window changed it in another. Cloud video renders offer the clip lengths the chosen model actually supports, read from the live catalogue, rather than a fixed list that lies about what the model will accept.

The first launch installs the engine and ComfyUI for you

The three steps are the product's onboarding in miniature. Install: run the installer, and on first launch a wizard looks for an AI engine already running on your machine and installs one with a single click if there is none, then does the same for ComfyUI, which draws the images and video. Pick a model: the manager marks which models fit your hardware and downloads them in one click, works with the engine you already have, naming Ollama and LM Studio among others, and can link the models those tools already store instead of copying them. Create: chat, create, code and agent are four tabs in one window, and the promise is that you type in one, switch to the next and keep the same models. That linking behaviour is the detail with real consequences, since a model library measured in tens of gigabytes is not something an installer should duplicate behind your back.

Two Vite configs and a dependency security test

The manifest shows a project that has been through some incidents, and the file also keeps the from-source path in a collapsed section at the bottom:

bash
git clone https://github.com/PurpleDoubleD/locally-uncensored.git
cd locally-uncensored
npm install
npm run dev          # browser dev mode
npm run tauri build  # desktop binary

Two bootstrap scripts sit alongside it, a batch file on Windows and a shell script on Linux and macOS, which install Node, Git and Ollama for development mode. Beyond the usual build, lint, typecheck and test scripts there are two you do not see in a typical application. One runs a circular dependency check over the TypeScript sources and reports cycles, which in a codebase this size is a deliberate guard rather than a nicety. The other is a security test for dependencies, run with the Node test runner against a script in the scripts directory, which means the dependency audit is something the project tests rather than something someone remembers to run. There is a browser end-to-end suite driven by Playwright, unit tests on Vitest, and Node 22 as the floor. Two Vite configuration files sit at the root, the normal one and a second named for a memory proof, which is the shape a specific regression investigation leaves behind. The Tauri shell plugin in the dependency list is the mechanism by which the app starts the engines it manages.

Working notes kept in the repository root

The root of the repository reads like a project diary. Alongside the readme, changelog, licence and contributing guide there are documents with names that are clearly internal: a goal file for macOS local support, a file about deferring the Linux updater, a Windows end-to-end handoff document, a Windows agent system prompt in plain text, an audit coverage document, an experiment changelog, and a plan for onboarding a built-in engine. Two of those have since been partly answered by the release notes, which is the usual fate of a plan document kept next to the code. There are also directories for a mobile client, a codex API surface, a separate development server and third-party material. The manifest is one of the few publishable ones in this batch of work, with the private flag set to false, an author field, and keywords covering local AI, Ollama, ComfyUI and the chat and image and video features. The example environment file holds a single commented variable, the path to a ComfyUI installation.

Editorial conclusion

This suits someone who already runs a local engine and wants one window instead of four tools, and the linking of existing model directories rather than copying them is the detail that decides whether the disk survives. Three things to know. Platform support is uneven in a specific way: Windows is described as tested on every release, Linux as built on every release, and macOS has no desktop build at all, with a browser-based cloud edition offered instead. The agent features are recent, since the most substantial fixes for the coding agent landed in the current release, four days after the last commit on master. And the licence is the copyleft one with the only- variant spelled out in the manifest, which matters if you intend to redistribute a build. The repository also keeps its working notes in the root rather than in a wiki.

Frequently asked questions

Does Locally Uncensored have a macOS build?

No. The download table lists macOS as having no desktop build and points to the cloud edition, which runs in the browser instead. Windows 10 and 11 get an installer, with an NSIS build recommended over an MSI, and Linux gets an AppImage, a deb and an rpm package.

Will antivirus software flag the Locally Uncensored installer?

Some engines do, because the installer is unsigned and downloads other binaries, which the file describes as a false positive. Its answer is verifiability: the installer is built by GitHub Actions from the public source on master, and the update channel is signed against a public minisign key, with both checks described in the repository's security document.

What does the coding agent in Locally Uncensored do?

It is one of four tabs, alongside chat, create and agent. The current release notes say writing a long file no longer ends the run, that continue picks up where a run stopped instead of starting over, that the agent keeps a plan only when you ask for one, and that cloud context defaults to 32K and can be set per model.

Which image and video models run locally in Locally Uncensored?

Qwen-Image 2.1 runs locally on Windows and Linux, generating an image from a prompt and editing one from a reference image with no mask to paint. It needs about 16 GB of files and ComfyUI 0.37.0 or newer, and carries a research licence that does not allow commercial use. After the previous release the cloud catalogue counted fourteen video models and seven image models.

What does Locally Uncensored need before I can use it?

Nothing manual. On first launch a wizard looks for an AI engine already running on the machine and installs one with a single click if there is none, then does the same for ComfyUI, which handles images and video. The model manager works with engines you already have, naming Ollama and LM Studio, and can link models those tools already store instead of copying them.

Official sources

  1. License: AGPL-3.0
  2. Project website
  3. PurpleDoubleD/locally-uncensored on GitHub
  4. README
  5. Releases
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/purpledoubled-locally-uncensored.svg)](https://hysenlabs.com/projects/purpledoubled-locally-uncensored)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/purpledoubled-locally-uncensored.svg)](https://hysenlabs.com/projects/purpledoubled-locally-uncensored)

Was this useful?