# Unsloth Desktop connects locally run models to Claude Code and Codex > Source: > Published: 2026-08-11 17:38:56+00:00 Brothers [Daniel and Michael Han](https://unsloth.ai/about) released a native desktop app for Unsloth AI on Tuesday, packaging the startup's local model training and inference software for Mac, Windows and Linux. [https://x.com/UnslothAI/status/2087177146662072546](https://x.com/UnslothAI/status/2087177146662072546) [Unsloth AI (@UnslothAI)](https://x.com/UnslothAI) announced the release [in a thread on X](https://x.com/UnslothAI/status/2087177146662072546) on August 11th. The open-source app supports local text, image, video and audio models, including models distributed in GGUF format and models built for Apple's MLX framework. Downloads are available through the [Unsloth GitHub repository](https://github.com/unslothai/unsloth) and Unsloth's main site. Unsloth calls the product the first desktop app capable of both running and training models locally. That is a company claim. The more consequential feature is the app's attempt to put inference, fine-tuning, data preparation and agent connections inside one interface, reducing the terminal work and separate tools typically required to customize an open model on a personal computer. The release also lets developers connect Claude Code and OpenAI's Codex to locally hosted models. Unsloth exposes OpenAI- and Anthropic-compatible endpoints, allowing supported coding agents to call a model running on the user's own hardware. Developers can keep a cloud model as the primary agent and assign local models as subagents, according to the project's documentation. That setup gives Unsloth a role beyond model optimization. The app can become the local model server behind coding tools that were designed around paid cloud APIs. The tradeoff moves to the user's machine: model quality, latency and usable context depend on the selected model and available memory, while prompts and inference can remain local. Unsloth says its "self-healing" system improves tool-call accuracy by 50%. The figure is company-supplied, and the launch thread does not specify the models, tasks or baseline used for the comparison. The underlying feature attempts to repair malformed tool calls from smaller open models, which often struggle to emit the exact structured output required by software agents. ### From kernels to a consumer-facing installer The desktop launch extends work Unsloth began with [Unsloth Studio](https://unsloth.ai/blog), a browser-based interface released on March 17th, 2026. Studio already handled local model downloads, inference, training, dataset creation and model exports. The new native packages make that system installable through standard downloads for Windows, macOS and multiple Linux formats, rather than requiring users to begin with shell or PowerShell commands. Daniel and Michael Han founded the San Francisco company in 2023 after building HyperLearn, an open-source library focused on faster machine-learning algorithms. Daniel previously worked at Nvidia and has described finding and fixing bugs in open models including Gemma, Llama, Mistral and Phi. Michael has focused on product, design and engineering. Unsloth joined [Y Combinator's Summer 2024 batch](https://www.ycombinator.com/companies/unsloth-ai), and [Lightspeed Venture Partners](https://lsvp.com/company/unsloth/) lists Unsloth as a portfolio company. The brothers initially built Unsloth around custom kernels that reduced the memory and time needed to fine-tune large language models. Their first public release in December 2023 was aimed at developers working through Python packages and notebooks. The desktop app carries the same optimization work into a product intended to be installed and operated without assembling a local training stack by hand. Unsloth's repository had roughly 69,900 GitHub stars on August 11th, 2026. The code uses a dual-license structure: the core package is available under Apache 2.0, while optional components including the Studio interface use AGPL-3.0. The current beta supports text, audio, embedding and vision models. Unsloth says Mac users can run MLX and GGUF inference and train models locally, while Windows and Linux users can train on supported Nvidia and AMD hardware. CPU-only systems are supported for narrower functions including chat and data preparation. The release puts Unsloth in direct competition with the collection of local model launchers, training frameworks and agent servers developers already combine themselves. Its advantage rests on integration: one installer now covers model discovery, training, inference, exporting, tool use and connections to coding agents. Whether that bundle holds together across the range of consumer hardware Unsloth supports will determine whether the desktop app broadens local model training beyond developers already comfortable managing Python environments and GPU dependencies.