Run private AI on your Mac
CereVault Serve
A lightweight Mac menu bar server for running and sharing local AI models across your trusted network. Built on Apple's MLX framework. Nothing leaves your devices unless you ask it to.
Screen recording of the CereVault Serve browser chat: a question is typed and sent, and the answer streams back from the model running on the Mac.
Chat from your iPhone or iPad.
Open the built-in web chat from any browser on your network.
Screen recording of the browser chat on a phone: a question about planning three days in Lisbon is sent and the answer streams in.
Swap models in a click.
Download once, activate any time. Everything stays on this Mac.
Screen recording of the Manage Models window: Activate is clicked on a downloaded model and the Active badge moves to it.
Share it only if you want to.
Keep it on this Mac, or let your devices in with an access code.
The Serve settings window with the network sharing options and the access code section highlighted.
One click in the menu bar.
Copy the server address, or jump to models and settings.
Your Mac becomes a personal AI server.
Bring your own model
Add any MLX model from Hugging Face. Serve reads its config first, checks the architecture will run, estimates the download size, and detects vision support before anything downloads.
Tool calling
The API accepts OpenAI-style tools and returns structured tool calls. Serve never runs a tool itself; your client does.
Curated model catalog
Pick from Llama 3.2, Gemma 3 and 4, Qwen3.5, and Phi 4 Mini with one click, including vision models. Each model downloads directly to your Mac.
Lightweight menu bar app
Lives quietly in your menu bar, with no Dock icon and no windows in your way until you need them.
Private by default
Models run entirely on your Mac. Network access is loopback-only until you explicitly turn on device sharing.
Local network sharing
Flip one toggle to let your iPhone, iPad, and other Macs on your Wi-Fi connect through CereVault Serve.
OpenAI-compatible API
Chat, streaming, vision, tool calling, and optional embeddings work with any client that supports a custom base URL.
Storage you control
Manage downloaded models in one place and free disk space anytime, typically 1–3 GB per model.
Optional access code
Require a code before anyone on your network can connect.
Browser chat
Anyone you share with can chat from a web page. Conversations are saved only on their own device.
Easy to find
While sharing is on, your Mac announces itself on your network with Bonjour so compatible apps can discover it. Nothing is advertised in loopback-only mode.
Embeddings, optional
Turn on an OpenAI-compatible embeddings endpoint in Settings. It is off by default and runs a small on-device model.
Models from a public registry, served locally.
When you choose to download a model, the files come directly from Hugging Face, a public model registry, and that connection is between your Mac and Hugging Face only. After that, everything runs on your machine. No accounts, no subscriptions, no analytics, and no data sent to the developer or any third party.
Requirements
- Apple Silicon Mac (M1 or later)
- macOS 14 or later
- 1–3 GB free disk space per downloaded model
- Wi-Fi or local network for device sharing (optional)
Get CereVault Serve
Free until Nov 1. Turn your Mac into a private, local AI server in minutes.
