Quail
Quick AI, local. A Postgres.app-style menu bar app for running local LLM servers on Apple Silicon.
Website and docs: quail-ai.app
Quail starts a runtime, exposes an OpenAI-compatible endpoint, and gets out of the way. It runs Quail server, its own server for both GGUF and MLX models, with llama.cpp's llama-server bundled as a fallback for GGUF, from one shared model folder and one honest opinion about what fits on your machine.
Install
Quail is free and open source, built for developers on Apple Silicon Macs. Homebrew is the recommended installation path; Mac App Store publication is not planned (ADR-0053).
brew install --cask adatoo/tap/quail-ai
or download the DMG from the latest release. The build is signed and notarized. Quail then updates itself (Quail → General → Updates); Homebrew leaves that to Quail unless you run brew upgrade --greedy. The brew install also puts the quail command on your PATH.
What it does
- Runs a model server from the menu bar. Start and stop it, see what it's doing, and keep the Mac awake while it works, even with a laptop's lid closed when plugged in.
- One server for GGUF and MLX. Quail server runs llama.cpp for GGUF and Apple's MLX for MLX models, from one model folder. llama.cpp's
llama-server stays bundled as a fallback for GGUF.
- Knows what fits. Every model gets a verdict for this Mac before you download it, with an estimated speed. Each model's context size and KV cache are labelled with how they fit.
- Finds models: a curated catalog, recommendations for your memory, over a hundred MLX models, any Hugging Face GGUF repo, and the models other apps already downloaded.
- Connects your tools. Copy-ready setup for Claude Code, Codex, opencode, Continue, Cline, Zed, Open WebUI and more, each with a Test. It speaks OpenAI Chat Completions, Responses and Embeddings, and Anthropic Messages, and reranks search results too.
- Includes a
quail command, like ollama: quail pull, chat, launch claude, bench, ps, logs and more.
- Benchmarks a model on your own Mac, the same way every time, and compares runs.
- Keeps your data on your Mac. It listens only to this Mac by default, requires an API key from the start, and works offline once models are downloaded (
task check:offline proves it).
What it deliberately doesn't do
It has no chat window of its own: the server's chat page opens in your browser, and quail chat works in the terminal. There are no agents, no image or audio generation, and nothing in the cloud: no account and no telemetry.
Status
1.0. What's planned next is at the end of docs/IMPLEMENTATION_PLAN.md: serving speech, images and MCP through the API (#179). Every merge to main is a signed, notarized release that installed copies update to.
Why Quail exists, and when Ollama or oMLX might suit you better: the launch post (source in docs/launch/).
Documents
| Doc | What it covers |
|---|
| docs/ARCHITECTURE.md | The design: distribution, app shell, runtime adapters, model store, device fit, packaging |
| docs/IMPLEMENTATION_PLAN.md | Phases, tasks, acceptance criteria, build/verify loop |
| docs/adr/ | Decision records with the trade-offs behind them |
| docs/benchmarks/ | Quail measured against llama-server, Ollama, oMLX and Rapid-MLX on the same weights, including where Quail is slower (ADR-0063). The first run is kept for comparison |
| AGENTS.md | Conventions for humans and coding agents working in this repo |
| website/ | The website at quail-ai.app: landing page and user docs, plain HTML (ADR-0062) |
Requirements
- macOS 14 Sonoma or later, Apple Silicon
- Xcode 16 or later to build
- An Apple Developer ID for signed/notarized builds (unsigned debug builds run locally)
Licence
MIT. The open-source components Quail ships, and their licences, are listed in THIRD_PARTY_NOTICES.md and in the app (About → Third-party licences).
