Three models, one API, all on your Mac

LocalLM Lab 1.0 adds a model layer: offer Apple's on-device model, Claude, and open-weight models you download and run locally (via MLX) behind one interface — routing, residency, and download/memory state the SDK owns, so you don't need to create your own provider abstraction. It comes with runnable example apps to build from, from a 100-line CLI to a sandboxed GUI.

⬇ Download the app Apple Silicon · signed & notarized · verify checksum

macOS 26 or 27 on Apple Silicon, Apple Intelligence enabled. On macOS 26 the app runs the Apple on-device model; Claude, Private Cloud Compute, and open-weight models need macOS 27. Building on the SDK? Start here — the SDK ships as binary xcframeworks; the reference apps and CLI toolkit are in the repo.

The model layer new in 1.0

Four built-in providers — SystemModelProvider (Apple on-device), ClaudeModelProvider, MLXModelProvider (downloaded open-weight models), and PCCModelProvider (Private Cloud Compute — not working in this beta) — behind one ModelProvider protocol. Name a model with a route (.heavy, .light, …), pick one per session, keep one warm between turns, and show download and memory state in your UI. It's opt-in and sits alongside the MCP client and connectors that have been here since 0.7 — which now run under any provider, so your MCP servers and connector tools are called the same way whether the reasoning runs on Apple's on-device model, Claude, or a downloaded open-weight model. The whole layer builds for macOS 26 — register SystemModelProvider there and the rest behind #available(macOS 27, *).

What the SDK gives you, and how to adopt it →
Which downloaded open-weight models actually tool-call — a 15-model field report →

Reference apps, not pseudocode

Eight small reference apps ship with the beta — each one actually runnable and documented, meant as a starting reference. For running a downloaded open-weight model there are three, smallest to fullest: repo-qa-local (one file, a CLI), code-buddy (a full local coding agent — two models, routing, tools, streaming, one-shot or an interactive loop), and workspace-buddy-local (a sandboxed GUI app, the Mac App Store shape). Each has a README with copy-paste setup and a real transcript.

Pick one to start from →

One Dock app, five panels

The menu-bar helper and its background Go server are gone. 1.0 is a single Dock app hosting Prompt Playground, Connectors, MCP Servers, API Lab, and AI Models (download, validate, and switch open-weight models) in one process — hotkey between them.

API Lab, rebuilt native

The OpenAI-compatible /v1 endpoint is now a native Swift HTTP server running in the app itself — localhost plus optional LAN HTTPS with a self-signed certificate — no separate server process. Point any OpenAI-SDK tool at your Mac, now backed by whichever model you've routed to.

Coming from 0.8.x?

The jump is mostly additive — the MCP client, connectors, and Keychain storage are unchanged, and the deployment floor stays at macOS 26. What can break a build: one enum-resilience change (a few switches need @unknown default), and — only if you wrote a custom ModelProvider — the protocol's makeSession change. See migrating-to-1.0.md and the beta notes.

Feedback

This is a beta — signatures can still move between releases, and a few things are rough (see the beta notes). Tell us what breaks: neuron@thisbrain.ai or the Discord.