Locally Uncensored
The all-in-one local AI studio: chat, image and video generation and a coding agent in one free, open source desktop app — no Docker, terminal or cloud required.
Evidence shows: agent tool calls pass a user-controlled permission gate with a read-only mode, coding agent shows diffs before applying, telemetry is limited to one anonymous daily count only when the Cloud switch is on and named in Settings, CI pins third-party actions to full commit SHAs, and SECURITY.md openly documents that the installer downloads and executes external binaries (Ollama, ComfyUI) with a minisign-signed update channel. Deductions: the permission gate, diff review, and read-only guarantees are README assertions with no implementation visible in the provided files; the agent's footprint (shell, file I/O, MCP, cloud lanes) is broad, and first-run auto-download-and-execute of other binaries is an inherently high-risk pattern; rollback beyond the signed update channel is unexplained.
Evidence shows: package. version 2.6.9 matches the README; CI gates lint, typecheck, import cycles, unit tests, e2e (Tauri-mocked), cargo test, clippy -D warnings, and a no-secret tauri-build job, with comments recording concrete past defects and their fixes (e.g. the v2.6.7 bundling failure); a failed build blocks release announcements (the 2026-08-15 incident). Deductions: distribution depends on GitHub Releases and downloading external engines at runtime; several CI claims (e.g. 0 lint errors) cannot be independently verified in this static review.
Evidence shows: aimed at desktop end users on Windows/Linux, macOS source-build only; the FAQ gives RAM/VRAM-tiered hardware guidance and a model manager that flags hardware fit; the roadmap and FAQ admit that Upscale/Erase Object are cloud-only and that frontier cloud models still win on hard reasoning. Deductions: trigger precision is weakly evidenced for a desktop agent; there is no macOS release, and wide multi-engine detection is asserted rather than demonstrated.
Evidence shows: a well-structured README (download, three steps, comparison, models, FAQ, roadmap, tech stack, community, license), a complete AGPL-3.0-only LICENSE, versioned releases with a CHANGELOG reference, a SECURITY.md with response timelines and scope, plus contribution guides and issue templates. Deductions: install and onboarding depth leans on the external locallyuncensored.com guides; the CHANGELOG file itself is not in evidence; the missing macOS release is an acknowledged gap.
Evidence shows: a free local all-in-one app combining chat, image/video generation and a coding agent, with gallery, side-by-side model benchmarking, and an OpenAI-compatible local API; hardware/model-size guidance lowers trial cost. Deductions: the comparison table is entirely self-assessed and lacks sourcing; real usability requires downloading 5–24 GB models, and output quality or agent effectiveness cannot be verified statically.
Evidence shows: key claims trace to file paths (install.rs, ollama.rs, release.yml, secret.rs:173), a community VirusTotal scan (Discussion #25), and a published minisign key; package., the CI workflows, and the README version agree. Deductions: many README assertions ('tested every release', 'no telemetry', 'every tool call passes a permission gate', 'a read only run stays read only') come with no verifiable evidence in scope, and factual claims are not clearly separated from promotional language; several dates are future-dated (September 2026) and can only be accepted as text.
- Security-critical claims (permission gate, diff review, read-only runs) are README assertions with no corresponding implementation visible in the reviewed files; verify them in source before relying on them.
- The installer is not Authenticode-signed and downloads and executes external binaries (Ollama, ComfyUI) on first run; although the update channel is minisign-signed, the first install is the weak link — download only from official Releases and check SHA-256 hashes.
- Some image lanes (Upscale, Erase Object) run in the cloud only, and the Cloud switch sends an anonymous daily count and can route data to hosted GPUs; privacy-sensitive users should take care.
- Model files range from 5 to 24 GB and hardware guidance is documentation only; validate VRAM/RAM requirements on a small scale first.
- Publisher identity is unverified (not in the FollowAgents registry), and 'uncensored/abliterated' models have refusal behaviour removed from weights — use within legal and compliance boundaries.
What does this agent do, and when should you use it?
Locally Uncensored is a free, AGPL-3.0 licensed desktop AI studio for Windows and Linux, built on Tauri v2 with a Rust backend, React 19 and TypeScript. It puts four capabilities in one window: chat (including refusal-free abliterated models), image and video generation through an app-managed ComfyUI, a coding agent that shows diffs before applying changes, and a general agent mode with tools and MCP servers. The app auto-detects existing engines (Ollama, LM Studio, llama.cpp and over a dozen others) or installs one with a single click, and the model manager flags which models fit your hardware before downloading. Everything runs on your own machine by default, with no accounts and no telemetry. Remote phone access works over LAN or a Cloudflare tunnel, and an optional Cloud switch (LU Labs Cloud) runs oversized models on hosted GPUs.
On first launch, a wizard detects an AI engine already running on your machine or installs Ollama with one click, and installs, starts, repairs and updates ComfyUI for image and video. The model manager filters models by hardware and downloads them in one click, linking models already stored by existing engines without copying. The Chat tab shows thinking as it happens, supports image understanding, document chat via a local index, voice in and out, cross-conversation memory, personas, and imports from ChatGPT/Claude/Gemini exports. The Create tab offers text-to-image, masked inpainting, background removal, image-to-video, animation, motion control, talking character, music (ACE Step 1.5 Turbo) and Character Studio (which trains a LoRA). The Code tab's coding agent builds a repo map, edits only requested lines, shows the diff before applying, runs tests and reads failures, honors per-project .lurules rules with Ask/Plan/Bypass modes; the Local API in Settings exposes all local models behind one OpenAI-compatible address with a token. The Agent tab adds web search/fetch, file read/write, a shell, code execution, screenshots, image/video generation and custom MCP servers, with background agents for long jobs and a permission gate on every tool call. It can also compare two models side by side and benchmark them on your hardware, with updates over a signed channel.
- Privacy-sensitive users who want refusal-free local models for chat and writing entirely offline — a 3B model runs on 8 GB RAM with no GPU
- Designers and content creators who want image and video generation on their own GPU without learning ComfyUI node graphs — 6-8 GB VRAM to start
- Indie developers who want a coding agent that edits code behind a diff review, runs tests, and follows per-project rules in a .lurules file
- Engineers who want their other coding tools to reach all local models through one OpenAI-compatible, token-protected endpoint
- Existing Ollama or LM Studio users who want one unified chat/image/agent UI — the app links models already stored by those tools
- People who want phone access to their desktop AI over LAN or a Cloudflare tunnel without data passing through third-party servers
What are this agent's strengths and limitations?
- Four-in-one: chat, image/video generation, a coding agent and a general tool-using agent in one free open source app — the README comparison table shows Open WebUI, LM Studio, Jan and SillyTavern each missing at least two of these
- No Docker, terminal or config files: the wizard auto-installs and maintains Ollama and ComfyUI, and the model manager marks hardware fit before download
- Detects and reuses models from a dozen-plus existing engines (Ollama, LM Studio, vLLM, KoboldCpp, llama.cpp, etc.) without copying files
- The coding agent shows the diff before applying edits, runs tests and reads failures, and the Local API offers a token-protected OpenAI-compatible endpoint for other tools
- Local mode has no accounts and no telemetry; remote access is off by default with QR/passcode pairing, and updates ride a minisign-signed channel
- No prebuilt macOS release — you must build from source with npm run tauri build, and a proper macOS build is still on the roadmap
- Some capabilities are cloud-only today: Upscale and Erase Object run through LU Labs Cloud; GLM 5.3's smallest local quant is 217 GB, out of desktop reach
- Some antivirus engines false-positive the unsigned NSIS installer that downloads other binaries; verification via SECURITY.md is on you
- Real hardware thresholds: image/video generation needs 6-8 GB VRAM minimum, and the Wan and HunyuanVideo models want 12 GB or more
- Abliterated models have refusal behavior stripped from the weights, which may not fit organizations that need policy-enforced guardrails
How do you install or deploy this agent?
- Grab the latest build from GitHub Releases (https://github.com/PurpleDoubleD/locally-uncensored/releases/latest): Windows 10/11 uses the .exe (NSIS, recommended) or .msi; Linux uses .deb, .rpm or .AppImage. 2. Run the installer like any normal program. 3. The first-launch wizard detects any existing engine (Ollama, LM Studio, etc.) or installs one with a single click; ComfyUI is installed automatically the same way. No macOS build is published yet — build from source: git clone https://github.com/PurpleDoubleD/locally-uncensored.git && cd locally-uncensored && npm install && npm run tauri build; setup.bat (Windows) or setup.sh (Linux/macOS) bootstraps Node, Git and Ollama for dev mode.
How do you use this agent?
After installing: 1. Pick a model in the model manager and download it in one click (e.g. Llama 3.1 8B Abliterated at ~5 GB; an 8B model runs comfortably on a 6 GB card). 2. Switch between the four tabs in one window: Chat for conversation, Create for images/video, Code to let the coding agent edit with diff review, Agent to delegate tool-using tasks behind permission gates. 3. To feed other tools, enable the Local API in Settings for a token-protected OpenAI-compatible endpoint. 4. Enable phone pairing (QR code plus passcode) for remote access. 5. Flip the Cloud switch for models too large for your card via LU Labs Cloud.
How does this agent compare with similar options?
The README compares it line by line against Open WebUI, LM Studio, Jan and SillyTavern: it alone offers image generation, video generation, image-to-video, a coding agent, a tool/MCP agent, model comparison and benchmarking, and out-of-the-box setup. LM Studio and Jan are also Docker-free and out-of-the-box but lack image/video/agent features; Open WebUI and SillyTavern need Docker or more complex setup; SillyTavern only partially supports refusal-free answers. The project site adds dedicated comparisons with GPT4All, Msty and KoboldCpp.