LU Labs (Locally Uncensored)

View Poster Visit Website Search Youtube Videos

LU Labs (Locally Uncensored)

LU Labs — the local AI studio that packs chat, code, images and video into one desktop app

LU Labs is the hosted-service side of Locally Uncensored (AGPL-3.0, by PurpleDoubleD), a free and open-source desktop app that runs open-weight models on your own machine. The pitch is simple: one window, four studios — Chat, Create, Code and Agent — instead of four separate tools, four model managers and four terminal windows. There is no Docker, no command line, no config file, and no account required to run locally. The desktop app has passed 170,000+ downloads and the GitHub repository sits at roughly 2.5k stars with about 400 forks.

The architecture is worth understanding because it explains the differences below: there is no server tier in the middle. On launch the app auto-detects the twelve local runtimes it supports — Ollama, LM Studio, Jan, GPT4All, KoboldCpp, llama.cpp, vLLM, LocalAI, text-generation-webui, TabbyAPI, Aphrodite and SGLang — and simply lists what your hardware can run. If you already run one of them, LU Labs uses it instead of installing a second copy of the same engine. If you run none, it ships its own LU Engine and installs one for you in a single click. It then installs, starts, repairs and updates ComfyUI for image and video work. In local mode every token, pixel and frame is produced on your own hardware; nothing is uploaded, nothing is logged, and there is no telemetry.

What is actually inside the four studios

Chat talks to any model your backend exposes, with thinking mode that shows reasoning as it happens, RAG over your own files and folders, voice input and read-aloud, personas, cross-conversation memory, and import of your existing ChatGPT, Claude or Gemini exports. Long threads auto-compact into a summary instead of running out of context. The model picker marks which models fit your VRAM before you download anything.
Create drives ComfyUI without exposing node graphs: text-to-image, image editing, inpainting with a mask, background removal, text-to-video, image-to-video, clip extension, motion control, talking character, music, and a Character Studio that trains a LoRA on your own card. FLUX.1 Schnell, FLUX 2 Klein, Z-Image Turbo, Qwen-Image 2.1, Wan 2.1/2.2, HunyuanVideo 1.5, LTX 2.3, AnimateDiff Lightning and FramePack F1 are all listed with their VRAM requirements.
Code is a coding agent that maps the repository, edits only the lines you asked for, shows the diff before applying it, runs your tests, reads the failures, and honours typed git tools plus per-project rules in a .lurules file. Ask, Plan and Bypass are per-conversation modes. A Local API in Settings exposes every local model on one OpenAI-compatible address behind a token, so your other coding tools can reuse the same machine.
Agent adds web search and fetch, file read/write, a shell, code execution, screenshots, image and video generation and your own MCP servers, with a permission gate on every tool call and long jobs pushed to background agents. A Computer Control mode (beta) can drive mouse and keyboard in apps you approve — it refuses to touch terminals, password managers or password fields, and every action that sends, buys or deletes needs your click.

Two smaller pieces round it out: Remote access from your phone over the LAN or a Cloudflare tunnel, paired by QR code and a six-digit passcode, and a compare-and-benchmark panel that runs two models side by side and measures speed, cost and correctness on your own hardware.

How it compares with the other local AI chat programs

This is the table the project publishes for itself. It is a vendor comparison, so read the individual claims skeptically — but the feature matrix is accurate against each product’s own documentation, and it lines up with what each tool actually ships.

Capability	LU Labs	Open WebUI	LM Studio	Jan	SillyTavern
Chat	Yes	Yes	Yes	Yes	Yes
Image generation	Yes	No	No	No	Partial
Video generation	Yes	No	No	No	No
Image-to-image and image-to-video	Yes	No	No	No	No
Coding agent	Yes	No	No	No	No
Agent with tools and MCP	Yes	No	No	No	No
Works out of the box	Yes	No	Yes	Yes	No
Remote access from your phone	Yes	Partial	No	No	Partial
Compare and benchmark models	Yes	No	No	No	No
Answers directly, without refusals	Yes	No	No	No	Partial
Voice in and out	Yes	Partial	No	No	Partial
Document chat	Yes	Yes	No	No	No
Runs without Docker	Yes	No	Yes	Yes	Yes
Open source	Yes	Yes	No	Yes	Yes

Against LM Studio. LM Studio is the closest competitor and the better-polished one for pure model management: a clean desktop GUI, an excellent model discovery and download experience, and a local OpenAI-compatible server. It is not open source, and it does not generate images, video, or run a coding agent. LU Labs gives up the polish and the model-download UX to gain a fourth studio, a diff-reviewing coding agent, a permission-gated tool agent, ComfyUI automation and an open-source license. If you only want to talk to one model and nothing else, LM Studio is the better tool; if you want one app instead of five, LU Labs is.

Against Jan. Jan is open source, ships its own engine, and is a genuinely good minimal local chat client with an extension system. Like LM Studio, it stops at chat. LU Labs is the same idea taken three studios further, and adds the auto-detection of twelve third-party backends so you are not locked to its own runtime.

Against Open WebUI. Open WebUI is the most common Ollama front end and is a fine server-side choice for a shared box. It is web-server shaped: typically Docker, usually pointed at an Ollama API, and it covers chat and document RAG well but not image or video generation and not a coding agent. LU Labs is desktop-app shaped — no server, no Docker, and the image and video path goes through ComfyUI on your own GPU.

Against GPT4All, Msty, KoboldCpp and text-generation-webui. These are single-purpose: a CPU-friendly chat runner (GPT4All), a polished chat-plus-knowledge-stacks client (Msty), a llama.cpp backend with a roleplay UI (KoboldCpp), or a very wide model-and-backend web UI (text-generation-webui). None of them generate images or video, and none of them edit your repository. LU Labs is the only one of the group where the chat window, the image canvas and the code editor are the same application sharing the same models. Ollama is the engine underneath most of them rather than a front end, and LU Labs detects and reuses it instead of asking you to run a second copy of the same model server.

Against SillyTavern. SillyTavern is excellent for roleplay and character cards, and LU Labs can import that style of workflow, but SillyTavern is built around a single use case. LU Labs is built around everything else, with roleplay as one of its modes rather than its purpose.

What it does better than the alternatives

One app instead of a stack. Chat, image, video, code and agent share one window and one model picker. Every competitor above requires a second app for images and a third for coding. LU Labs also auto-detects and reuses twelve existing runtimes, so it is often an addition to your setup rather than a replacement for it.
Real media generation, not a wrapper. ComfyUI is driven for you: installed, started, repaired and updated automatically, with FLUX, Qwen-Image, Wan, HunyuanVideo, LTX, ACE-Step music and LoRA training on your own card. Unlimited local generations, no watermark, and no per-image billing.
A coding agent you can audit. It shows the diff before applying it, runs the tests, reads the failures, and gives you Ask / Plan / Bypass modes per conversation. That diff-first review step is what separates it from a chat box that claims to edit code.
Permissions, not just a sandbox claim. Every agent tool call passes a permission gate you control; a read-only run stays read-only; and Computer Control never touches terminals, password managers or password fields.
Genuine privacy in local mode. No account, no analytics SDK, no crash pings, no cloud credentials to leak, and it keeps working with the ethernet cable pulled. The ComfyUI process is killed when the app quits, so nothing keeps listening. Your chats live on your own disk.
An escape hatch when your card is too small. Flip the Cloud switch and the same app and the same account run 47 chat models, 7 image models and 5 video models on managed NVIDIA H100, A100 and B200 hardware, in the browser on a phone or a Mac with no GPU. The catalog is open-weight only, the data is hosted in the EU, and the provider states it never trains on your data and never sells it.

Honest limits

Hardware still decides the ceiling. Chat runs a 3B model on 8 GB of system RAM with no GPU at all, and an 8B model comfortably on a 6 GB card. Image generation starts at 6–8 GB of VRAM, and Wan and HunyuanVideo want 12 GB or more. GLM 5.3’s smallest local quant is 217 GB, so anything at that scale is a cloud job by definition, and Kimi K3 is further out still.
macOS is a beta. The Mac build needs Apple Silicon and macOS 14 or newer; Intel Macs are unsupported. On a Mac, chat runs through the bundled server with Metal and Create makes images and video natively without ComfyUI, so the ComfyUI and VRAM guidance describes Windows and Linux. Remote access and automatic update are not confirmed on macOS.
Some features are still beta. Voice is labelled beta and was checked by hand on a Mac but on Windows in the real app; Computer Control was verified on Windows only and is explicitly not proven in every app. Create in chat is also beta.
Two Create lanes are cloud-only. Upscale and Erase Object run on hosted GPUs today, not locally, so a fully offline image workflow is not quite complete.
Unsigned Windows installers, for now. The production SignPath code-signing certificate is not in place yet, and some antivirus engines flag unsigned NSIS installers that fetch other binaries. The project documents this as a false positive and points to its signed minisign update channel, but enterprise machines with strict allow-listing will still object.
AGPL-3.0 is a strong copyleft license. Fine for personal and internal use; if you build a network service on top of it, the source obligations apply.
The uncensored angle is a real differentiator, and a real responsibility. Abliterated weights have the trained refusal behaviour removed from the weights themselves, so the model answers directly. The project states that minors-related content is refused everywhere, always, and that for hosted image and video the account’s own content policy applies — but you are the one responsible for what you run locally.
Cloud is paid and metered. Local is genuinely free forever. Hosted plans are credit-based (from €19/month for 900,000 credits, €49 Pro, €99 Max, or one-time credit packs from €15 that never expire) and the 11–12 Flash chat models have a separate daily allowance. There is no five-hour window, but there is also no unlimited hosted promise.

Who should pick it

Pick LU Labs if you want a private, offline, all-in-one studio and you are willing to deal with local hardware limits, or if you want one app that spans chat, image, video and code today. Pick LM Studio if model discovery and a clean chat UI are all you need and you do not mind a closed-source tool. Pick Jan or GPT4All if you want a lightweight open-source chat client. Pick Open WebUI if you are running a shared server for a team rather than a personal workstation. Pick SillyTavern if roleplay is the whole point.

For most people the deciding question is not features but GPUs: the local path is excellent and free, and the cloud path is a paid convenience for models no desktop card can hold.

AIToolScan