Brazier is a free, open-source desktop client and local API for running open weight AI models. Chat with LLMs, run computer use to operate your browser and desktop, run sandboxed agents, generate images and video, and build with a local OpenAI-compatible API — no cloud, no telemetry, all on your hardware.
See all releases · Latest: v0.2.10-beta
An observe–act loop for both the browser and your desktop. It drives isolated, headless Chromium sessions through CDP, and — when OS permissions allow — controls the desktop via macOS Accessibility/Screen Recording or Linux X11/Wayland portals, failing closed until capture and input are granted. Tasks, steps, and memories persist in the daemon, so an interrupted job can be resumed after a restart.
Run local language models with llama.cpp, MLX-LM, and MLX-VLM backends. Persistent branching conversations, search, attachments, cancellation, and import/export.
A broker-sandboxed agent core with its own tools, policy approvals, and OS sandboxing (Seatbelt on macOS, Bubblewrap on Linux). Pi provides the orchestration loop; optional Oh My Pi adds a fuller coding harness.
Generate media with stable-diffusion.cpp, and let LLMs call media generation as tools in multi-round conversations.
Build engine forks straight from a repository URL — llama.cpp, MLX, whisper.cpp, stable-diffusion.cpp, and PersonaPlex. When a new model lands in an upstream fork, you don't have to wait for a Brazier release.
An OpenAI-compatible API with Models, Chat Completions, and Responses, plus transcription, generation, and agent endpoints.
Managed runtimes, system runtime discovery, and activation. Hardware-sized model recommendations and Hugging Face search. No telemetry, ever.
MIT-licensed. Runs on macOS (Apple Silicon) and Linux, built with Electron.