A Windows desktop app called SEELS that allows users to run local LLMs, correct model responses, and automatically train LoRA adapters from corrections. It bundles voice mode, hardware dashboard, and a training sidecar.
i built a local AI desktop app and just shipped it. windows only. called SEELS. dropping it here cause if anyones gonna find the cracks its you guys. the thing i actually wanted to make wasnt another ollama wrapper. what bugged me is every local model id run would say something dumb and id sit there going "no thats not what i meant". then id close the chat and the model never knew, never learned. so the whole hook of SEELS is theres a Teach button on every reply where you write what it should have said. those corrections pile up into a jsonl corpus, and when you have enough you click Train and it actually kicks off a PEFT LoRA run on your base. no notebook, no python, no terminal. just chat, correct, train. over time the adapters stack up and it becomes your model not theirs. trained a tiny 0.6B helper from scratch on like 110 hand written examples so theres something that runs on CPU out of the box. not replacing your daily 35B obviously but it answers questions about how to use SEELS itself which was the point. rest of standard (free, forever, not a trial): bring any GGUF, voice mode with whisper STT and piper TTS both local no API keys, hardware dashboard so you can stop guessing what your card has free, single instance lock cause i kept opening two and corrupting my own sqlite, and 14 settings tabs because i couldnt cut any of them. wont lie. pro tier (image/video/music gen, code workspace, multi profile, multi lora stacking, MCP, cron) is written and gated behind a tier check but not purchasable yet. max tier (mask editor, plugin sandbox, comfyui backend, node graph workflows, multi GPU, seels-cli) is roadmap, some of those literally have no code yet. didnt want to put 12 things on a tier card if 8 of them are vibes. so the site says "ships in waves" instead of lying. things i actually need eyes on: if you teach it 30 things and the next train run is worse, thats a real bug, settings > training proof has a copy button for the trajectory log "failed to load model" used to be one useless message, now it branches into OOM / corrupt / unsupported arch / missing file. tell me which one fired and on what model if it lies to you big red banner now appears whenever voice mode is on cause someone yelled at me that the mic felt sneaky. tell me if its too much. installer is 2.81 GB, bundles cuda runtime + a portable python for the training sidecar. signed via azure trusted signing so no "unknown publisher" warning, but smartscreen will still yell "windows protected your pc" the first dozen installs cause brand new app. theres a heads-up modal that walks you thru More Info > Run Anyway. its 0.1.5 alpha. real alpha. things break. rather find em with people who know what theyre looking at. [tideforge.ai/seels](http://tideforge.ai/seels)
MLX-LoRA-Studio is a native macOS app for fine-tuning LLMs on Apple Silicon, offering a user-friendly interface and support for various training algorithms including SFT, DPO, and QAT. It is fully open-source and allows local, private fine-tuning without cloud dependency.
An open-source tool for training LLMs on consumer hardware, featuring real-time neural visualization for hallucination detection and model introspection, currently supporting small-scale models.
llama.cpp Console is a Windows desktop app that provides a GUI for managing llama.cpp in WSL/Ubuntu, handling installation, building, model downloading, and serving.
A fully offline, CPU-only voice loop for local LLMs using Silero VAD, Parakeet STT, and Supertonic TTS, integrated via a one-command installer. Works with Ollama, LM Studio, and various agent frameworks.