Locally Uncensored

An AGPL-3.0 desktop application that puts local AI chat, image generation, video generation and a coding agent behind one installer. Tauri v2 (Rust) with React 19 / TypeScript / Tailwind 4 / Vite 8; Windows 10/11 and Linux, macOS "coming soon". Repository metadata read 2026-09-04: 1,374 stars, 213 forks, 7 open issues, 1,317 commits, created 2026-03-24, last push 2026-09-03, latest release v2.6.7 (2026-08-31), by "LU Labs" under the handle PurpleDoubleD.

Its pitch is three refusals at once, and the README states each in a headline: no cloud ("100% local and private — zero telemetry, zero analytics, works fully offline after model download"), no setup ("no command line, no Docker, no config files"; a wizard auto-detects or installs any of twelve local backends), and no refusals from the model — first-class support for abliterated models, which it correctly describes as having "the refusal behavior removed from the weights themselves — not a jailbreak." The tagline: "Your AI, your hardware, your rules."

What it actually bundles

Chat — 20+ provider presets. Local: Ollama, LM Studio, vLLM, KoboldCpp, llama.cpp, LocalAI, Jan, TabbyAPI, GPT4All, Aphrodite, SGLang, TGI. Cloud, optional and bring-your-own-key: OpenAI, Anthropic, OpenRouter, Groq, Together, DeepSeek, Mistral. Thinking mode, vision upload, a memory system, RAG over local embeddings, voice (faster-whisper STT + neural TTS), 25+ personas, and import from ChatGPT/Claude/Gemini exports.
Create — a bundled, auto-managed ComfyUI. Images: FLUX 2 Klein, FLUX.1, Juggernaut XL, Z-Image Turbo, ERNIE-Image, SDXL, SD 3.5. Video: Wan 2.1/2.2, HunyuanVideo 1.5, LTX 2.3, AnimateDiff, Mochi, Cosmos, with image-to-video via FramePack F1 claimed at 6 GB VRAM.
Code — a coding agent with an Aider-style PageRank repo-map, a line-scoped file_edit tool, review-before-apply diffs and a test-runner loop; plus Agent Mode with 28 tools and MCP and a seven-category, three-level permission model.
Everywhere — a mobile web app reached over LAN or a Cloudflare Tunnel, with QR pairing and a six-digit passcode.

75+ one-click model downloads filtered by what the machine can run, from Llama 3.1 8B Abliterated at 6 GB — "the classic uncensored starting point" — up to GPT-OSS 120B via Ollama.

The claim and the second mode

The README's economic promise is unambiguous: "Free forever — AGPL-3.0. The models are open weights. No subscription, no credits, no rate limits." The FAQ repeats it: "no account, no telemetry, no usage limits."

The changelogs in the same file describe LU Cloud: a credits meter, per-second billing for music generation, cloud-only lanes (Upscale, Erase Object, Character Studio), and a cloud model catalogue with prices in the picker. Two of the fixes are about being charged:

2.6.4 — "The Cloud switch now really means off: flipping to Local with no local model running used to keep the cloud model silently active and kept billing credits; the send path now refuses any model from the wrong mode."

2.6.4 — the music price "followed the length slider" only after the fix: billing was per-second while the label quoted one minute, "so a 3 minute song looked three times cheaper than it was."

And 2.6.6 measures the paid path against itself: "Measured over our own set of tool driven runs, 2.6.6 spent 78.6 percent fewer credits than our own 2.6.5."

None of this is concealed — it is in the README, in plain words, under "What's new." But the headline and the changelog describe different products, and the failure mode in 2.6.4 is precisely the seam between them: a user who believed the headline, with the switch set to Local, was billed. This is the corpus's first case where the gap between the pitch and the product had a meter attached.

Distribution, and what the downloads say

Windows-first by a wide margin, and the release page proves it. On v2.6.7: the Windows NSIS installer took 8,563 downloads against 244 (AppImage) + 212 (.deb) + 120 (.rpm) for all of Linux — roughly fifteen to one — with the .msi at 123. Above both sits latest.json at 11,308, the manifest the signed auto-updater polls, which is a rough count of installs checking in rather than of new users.

There is no store anywhere in that path. The project ships bare installers from GitHub Releases, signed against "a public minisign key," and the README carries its own antivirus warning:

"Some engines flag unsigned NSIS installers that download other binaries — a false positive. The installer is built by GitHub Actions from the public source on master."

That is a real description of what the installer does: it fetches an inference engine and a ComfyUI environment at setup time. The project's answer to being flagged is a reproducible-build argument plus a signing key it publishes itself — self-attestation, with no third party in the loop. Set beside f-droid (a repository that signs on the developer's behalf) and termux-launcher (bare APKs, no store, no auto-update), this is a third position: no store, but a signed update channel the project runs.

How it describes its own quality

The 2.6.7 notes are unusually specific about method: "Every fix went back to a fresh tester who did not know what had been changed, driving the real installed build on Windows, and whatever they broke went into the next round; that loop ran seventeen times." First-party and unverifiable, like every process claim in this corpus, but it is a claim about how the fixes were validated rather than a count of them — and the fixes it lists (an engine held as an orphan process after exit, video decode paging a 12 GB card into an hour-long stall, an AppImage leaking its bundled libraries into every Python it launched) read like the output of exactly that loop.

The comparison table, and the frame

The README compares itself down a column of eleven rivals' worth of features against Open WebUI, LM Studio, Jan and SillyTavern, and links eight "vs" blog posts on its own domain. It is the open-source-alternative frame applied not to a proprietary incumbent but to other FOSS applications — the substitution move turned sideways. Two of its four comparators are AGPL like it is. What the table sells is not freedom from a vendor; it is consolidation, one installer replacing four.

Cross-spoke

The abliteration story belongs to ../llm-providers-wiki: the mechanism and its measured cost, the heretic tool that made the edit a commodity, and the finding that on open weights a model's safety posture is a default that ships with an uninstaller. This page is the consumer end of that. Locally Uncensored is the first source in this corpus that turns it into a menu item for a non-technical Windows user — hardware-filtered, one-click, with a beginner guide and a Discord. The supply side stopped being hard in 2024; this is what the demand side looks like two years later.

foss-application · self-hosting · open-source-alternative · f-droid · termux-launcher · pocketpal-ai · hometube · ceefax-station · abliteration · heretic