Complete Guide: Always-On Mac Server with MLX, MLX Manager & QwenPaw
This guide transforms your Apple Silicon Mac into a persistent, always-on local AI server. By the end, you'll have MLX models running natively on the Metal GPU, a web UI for model management, and the QwenPaw agent framework — all surviving reboots and running headlessly with secure remote access via MagicDNS.

Part 1: Prepare macOS for Headless Server Mode
These steps ensure your Mac never sleeps, is accessible remotely, and boots into a usable state without manual intervention.
1.1 Prevent System Sleep
Run this in Terminal to disable all sleep modes permanently:
sudo pmset -a disablesleep 1 sleep 0 disksleep 0 displaysleep 0 standby 0 autopoweroff 0 powernap 0
Expected output: You'll be prompted for your password. After entry, no confirmation message appears — that's normal.
Verify with:
pmset -g
Expected output:
System-wide power settings:
SleepDisabled 1
Currently in use:
standby 0
Sleep On Power Button 1
autorestartatconnect 0
autorestart 0
SleepServices 0
powernap 0
networkoversleep 0
disksleep 0
sleep 0 (sleep prevented by powerd, bluetoothd)
ttyskeepawake 1
displaysleep 0
tcpkeepalive 1
powermode 0
womp 1
Note: On macOS Ventura (13) and later,
pmsetreports the disablesleep setting asSleepDisabled 1instead ofdisablesleep 1. This is expected — it means the same thing.The annotation
(sleep prevented by powerd, bluetoothd)next tosleep 0is informational, not an error. It simply means that in addition to yourpmsetsettings, system daemons are also actively preventing sleep. Your server will stay awake.
1.2 Enable SSH (Remote Login)
On macOS Catalina (10.15) and later, the systemsetup command requires Full Disk Access to toggle Remote Login. Running it without that permission will fail with:
setremotelogin: Turning Remote Login on or off requires Full Disk Access privileges.
This is a macOS security feature (TCC), and sudo alone does not bypass it. Follow these steps to fix it permanently.
Step 1: Grant Full Disk Access to Terminal
- Open System Settings.
- Go to Privacy & Security → Full Disk Access.
- Click the + button.
- Navigate to Applications → Utilities and select Terminal.app (or your terminal of choice, e.g., iTerm).
- Toggle Terminal on in the list.
- Quit Terminal completely (
Cmd+Q) and reopen it. The permission only takes effect for new sessions.
Step 2: Enable Remote Login
Now run the command. It will succeed without the permission error:
sudo systemsetup -setremotelogin on
Expected output: setremotelogin: remote login is now on
Verify:
sudo systemsetup -getremotelogin
Expected output: Remote Login: On
Now you can SSH from another Mac: ssh username@your-mac-ip
1.3 Enable Auto-Login (Required for launchd Services at Boot)
launchd user agents only start after a user logs in. Without auto-login, your services won't run after a reboot.
- Open System Settings → Users & Groups
- Click Automatic Login → select your user account
- Enter your password when prompted
Note: FileVault must be OFF for auto-login to work. Go to System Settings → Privacy & Security → FileVault and turn it off if it's enabled. This is acceptable for a physically secure home server.
1.4 Optional: Menu Bar Server Mode Toggle
For quick control without terminal, install the mac-server-mode menu bar app:
git clone https://github.com/thairc-dev/mac-server-mode.git
cd mac-server-mode
./build.sh
nohup ./mac-server-mode >/dev/null 2>&1 &
./install-launch-agent.sh
A ⚡ icon appears in your menu bar. Press Ctrl + Option + Cmd + S to toggle server mode (screen off, system awake).
Part 2: Install uv & Python 3.11
2.1 Install uv
uv is a fast, all-in-one Python package and version manager written in Rust. Install it via the official one-liner:
curl -LsSf https://astral.sh/uv/install.sh | sh
Expected output: Installation progress bars, then a success message.
After installation, restart your shell or source your profile:
source ~/.zshrc
Verify:
uv --version
Expected output: uv 0.x.x
2.2 Install Python 3.11 with uv
MLX Manager requires Python 3.11 or 3.12. Install Python 3.11 as a uv-managed Python:
uv python install 3.11
Expected output: Download progress, then Installed Python 3.11.x
Verify:
uv python list | grep 3.11
Expected output: Shows the installed Python 3.11 version.
2.3 Create a Virtual Environment for MLX
Create a dedicated virtual environment using Python 3.11:
uv venv --python 3.11 ~/.mlx-venv
Activate it:
source ~/.mlx-venv/bin/activate
Verify the Python version inside the venv:
python --version
Expected output: Python 3.11.x
Critical: Ensure you are NOT using a Rosetta (x86) Python. Run this check:
python -c "import platform; print(platform.processor())"Expected output:
arm— if you seei386, you have the wrong Python.
Part 3: Install MLX Framework
MLX is Apple's native machine learning framework for Apple Silicon.
With your virtual environment activated:
uv pip install mlx mlx-lm
Expected output: Successfully installed mlx-x.x.x mlx-lm-x.x.x ...
Verify:
python -c "import mlx; print(mlx.__version__)"
Expected output: 0.x.x
Part 4: Install MLX Manager (Web UI + Background Service)
MLX Manager provides a browser-based interface for downloading MLX models from Hugging Face and running them as persistent services. It is distributed via a Homebrew tap, so we need to install Homebrew first.
4.1 Install Homebrew
Homebrew is the missing package manager for macOS. Install it with the official one-liner:
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
Expected output: Installation progress messages, followed by:
==> Installation successful!
==> Homebrew has enabled anonymous aggregate formulae and cask analytics.
...
==> Next steps:
- Run these commands in your terminal to add Homebrew to your PATH:
echo >> ~/.zprofile
echo 'eval "$(/opt/homebrew/bin/brew shellenv)"' >> ~/.zprofile
eval "$(/opt/homebrew/bin/brew shellenv)"
Follow the printed Next steps to add Homebrew to your PATH (Apple Silicon uses /opt/homebrew; Intel uses /usr/local):
echo >> ~/.zprofile
echo 'eval "$(/opt/homebrew/bin/brew shellenv)"' >> ~/.zprofile
eval "$(/opt/homebrew/bin/brew shellenv)"
Verify:
brew --version
Expected output: Homebrew 4.x.x
Note: If Homebrew is already installed, this command will detect it and skip installation. You can safely run it again.
4.2 Install MLX Manager via Homebrew
brew tap tumma72/mlx-manager https://github.com/tumma72/mlx-manager
brew install mlx-manager
Expected output: /opt/homebrew/Cellar/mlx-manager/x.x.x: ...
4.3 Start the Web UI
mlx-manager serve
Expected output:
INFO: Started server process [xxxxx]
INFO: Waiting for application startup.
INFO: Application startup complete.
INFO: Uvicorn running on http://0.0.0.0:8080
Open http://localhost:8080 in your browser (or http://your-mac-ip:8080 from another device).
4.4 Create Your Account
On first visit, you'll see a registration page. The first user automatically becomes admin. Create your account and log in.
4.5 Install as Persistent launchd Service
This is the key step for always-on operation. Stop the foreground server first (Ctrl+C), then:
mlx-manager install-service
Expected output:
✅ launchd service installed: com.mlx-manager.server
Service will auto-start on login.
Use `mlx-manager status` to check.
Verify it's running:
mlx-manager status
Expected output: Shows the service as running with a PID.
4.6 Download & Deploy a Model
- In the web UI, click Browse to search Hugging Face for MLX models (e.g.,
mlx-community/Qwen3.5-4B-8bit) - Click Download — watch the progress bar
- Click Configure to create a server profile (GPU layers, context length, etc.)
- Click Run to start serving
Once running, the model is available at:
http://localhost:10242/v1
Test it:
curl http://localhost:10242/v1/models
Expected output:
{"object":"list","data":[{"id":"mlx-community/Qwen3.5-4B-8bit","object":"model"}]}
Part 5: Install QwenPaw (Agent Framework)
QwenPaw is a local-first AI agent framework that can connect to your MLX-served models.
5.1 Install via One-Line Script
curl -fsSL https://qwenpaw.agentscope.io/install.sh | bash
Expected output:
✅ QwenPaw installed successfully to ~/.qwenpaw
✅ uv Python environment created
Run `qwenpaw init --defaults` to configure.
This script installs QwenPaw into ~/.qwenpaw with a self-contained uv-managed Python environment. It automatically installs uv, creates a virtual environment, and downloads dependencies without requiring manual Python setup.
5.2 Initialize QwenPaw
qwenpaw init --defaults
Expected output:
✅ Configuration initialized at ~/.qwenpaw/config.json
✅ Default model backend configured
5.3 Configure QwenPaw to Use MLX Manager
Edit the QwenPaw config to point to your MLX Manager API:
vim ~/.qwenpaw/config.json
Find the model section and set:
{
"model": {
"backend": "openai",
"base_url": "http://localhost:10242/v1",
"api_key": "not-needed",
"model_name": "mlx-community/Qwen3.5-4B-8bit"
}
}
5.4 Start QwenPaw
qwenpaw app
Expected output:
QwenPaw Console running at http://127.0.0.1:8088/
Open http://localhost:8088 to access the QwenPaw web console.
5.5 Make QwenPaw Persistent (launchd)
Create a launchd plist so QwenPaw starts automatically on login. The heredoc below is unquoted (<< EOF, not << 'EOF'), so the shell expands $HOME to your actual home directory when the file is written:
mkdir -p ~/.qwenpaw/logs
cat > ~/Library/LaunchAgents/com.qwenpaw.app.plist << EOF
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
<key>Label</key>
<string>com.qwenpaw.app</string>
<key>ProgramArguments</key>
<array>
<string>$HOME/.qwenpaw/bin/qwenpaw</string>
<string>app</string>
</array>
<key>RunAtLoad</key>
<true/>
<key>KeepAlive</key>
<true/>
<key>WorkingDirectory</key>
<string>$HOME</string>
<key>StandardOutPath</key>
<string>$HOME/.qwenpaw/logs/qwenpaw.log</string>
<key>StandardErrorPath</key>
<string>$HOME/.qwenpaw/logs/qwenpaw.error.log</string>
</dict>
</plist>
EOF
Verify the paths resolved correctly (they should show your real username, not $HOME):
grep -A1 ProgramArguments ~/Library/LaunchAgents/com.qwenpaw.app.plist
Expected output: The <string> lines should contain e.g. /Users/yourname/.qwenpaw/bin/qwenpaw.
Load it:
launchctl load ~/Library/LaunchAgents/com.qwenpaw.app.plist
Verify:
launchctl list | grep qwenpaw
Expected output: - 0 com.qwenpaw.app
Part 6: Verify Everything Survives a Reboot
Restart your Mac:
sudo reboot
After reboot, wait 30 seconds, then SSH in from another machine:
ssh your-username@your-mac-ip
Check all services:
# MLX Manager
mlx-manager status
# QwenPaw
launchctl list | grep qwenpaw
# MLX Manager web UI
curl -s http://localhost:8080/health
# QwenPaw console
curl -s http://localhost:8088/
All should return healthy responses without any manual intervention.
Part 7: Remote Access from Anywhere with Tailscale & MagicDNS
Tailscale creates a secure mesh VPN between your devices, giving each one a stable private IP address that works from any network — even behind CGNAT or corporate firewalls, with no port forwarding required.
MagicDNS is Tailscale's built-in DNS feature that lets you reach your server by its machine name (e.g., my-mac) instead of remembering IP addresses. This section sets up MagicDNS as the primary access method for all your services.
7.1 Install Tailscale on Your Mac Server
brew install --cask tailscale
Expected output: tailscale was successfully installed!
Open the Tailscale app from your Applications folder or menu bar, and sign in with Google, GitHub, Microsoft, or email.
Once authenticated, the Tailscale icon appears in your menu bar and shows a green "Connected" status.
Note: The
--caskversion provides the GUI app and includes thetailscaleCLI command automatically. For a headless server setup, this is the recommended approach.
7.2 Install Tailscale on Your Client Device
Install Tailscale on any device you'll use to connect — your laptop, phone, or tablet:
- macOS:
brew install --cask tailscaleor download from the Mac App Store - Windows: Download the installer from tailscale.com/download
- Linux:
curl -fsSL https://tailscale.com/install.sh | sh && sudo tailscale up - iOS/Android: Download from the App Store or Play Store
Critical: Sign in with the same account on every device. All devices must belong to the same "tailnet" (your private Tailscale network) to see each other.
7.3 Enable MagicDNS and HTTPS Certificates (One-Time Setup)
MagicDNS and HTTPS Certificates are required for secure, name-based access to your services.
- Go to the DNS page of the Tailscale admin console.
- Ensure MagicDNS is enabled. Tailnets created on or after October 20, 2022 have MagicDNS enabled by default.
- Under the HTTPS Certificates section, click Enable HTTPS. This allows Tailscale to automatically provision TLS certificates for your devices.
Important: Enabling HTTPS certificates publishes your machine names and tailnet DNS name on a public ledger (Certificate Transparency). If your machine names contain sensitive information, rename them before enabling HTTPS.
7.4 Find Your MagicDNS Name
Once MagicDNS is enabled, you can reach your Mac server by its machine name. There are two ways to find the name:
Method 1 — Menu Bar (Easiest): Click the Tailscale icon in your Mac's menu bar. Your machine name appears under "This Device".
Method 2 — Terminal:
tailscale status
Expected output:
100.101.102.10 my-mac your-email@example.com macOS -
100.101.102.11 my-laptop your-email@example.com macOS -
Your MagicDNS name is the second column (e.g., my-mac).
Method 3 — Admin Console: Go to console.tailscale.com/admin/machines, find your Mac in the list, and copy the machine name.
Tip: You can rename your device to something memorable (like
ai-server) by editing the machine name in the admin console. The MagicDNS entry will update automatically.
7.5 Verify MagicDNS Resolution
From your client device (e.g., your laptop), test that MagicDNS resolves the machine name:
ping my-mac
Expected output:
PING my-mac (100.101.102.10): 56 data bytes
64 bytes from 100.101.102.10: icmp_seq=0 ttl=64 time=12.3 ms
If you get a response, MagicDNS is working. Press Ctrl+C to stop.
Note: Some CLI tools on macOS like
hostornslookupbypass system DNS resolution and will not work with MagicDNS. Usepingorcurlinstead.
7.6 SSH into Your Mac Server Using MagicDNS
Once MagicDNS is working, SSH becomes much simpler — no more remembering IP addresses:
ssh your-username@my-mac
Expected output:
The authenticity of host 'my-mac (100.101.102.10)' can't be established.
ED25519 key fingerprint is SHA256:...
Are you sure you want to continue connecting (yes/no)?
Type yes and enter your Mac's login password. You are now remotely connected to your always-on AI server from any network in the world, using a human-readable name.
Tip: For shared devices, you must use the full domain name:
ssh your-username@my-mac.your-tailnet.ts.net.
7.7 Optional — Enable Tailscale SSH (Passwordless)
Tailscale SSH lets you authenticate using your Tailscale identity instead of your macOS password. This is useful for automation and scripting.
On your Mac server, run:
tailscale set --ssh
Expected output: No output (success).
Verify with:
tailscale status --json | grep -i ssh
Expected output: "ssh": true
Now when you SSH from another Tailscale device, Tailscale handles authentication automatically:
ssh your-username@my-mac
Important: Tailscale SSH requires you to configure an access control policy. Go to console.tailscale.com/admin/acls and add:
"ssh": [ { "action": "accept", "src": ["your-tailscale-username"], "dst": ["autogroup:self"], "users": ["autogroup:nonroot"] } ]This allows devices you own to SSH into each other without a password.
7.8 Secure Remote Access to Web UIs with HTTPS via MagicDNS
Your MLX Manager (port 8080) and QwenPaw (port 8088) web interfaces can be accessed securely over HTTPS using Tailscale Serve. This provides a valid TLS certificate automatically and encrypts all traffic.
With MagicDNS enabled, you access these services by name instead of IP. Tailscale Serve keeps your web UIs bound to localhost while giving your tailnet an HTTPS MagicDNS hostname.
Exposing your web UIs:
Run these commands on your Mac server:
tailscale serve --bg --https=443 http://localhost:8080
tailscale serve --bg --https=443 http://localhost:8088
Expected output:
Available within your tailnet:
https://my-mac.your-tailnet.ts.net/
|-- proxy http://localhost:8080
Tailscale will print the HTTPS MagicDNS URL for your services.
Accessing your services using MagicDNS:
Open the HTTPS URL provided by Tailscale in your browser from any device on your tailnet:
https://my-mac.your-tailnet.ts.net (for MLX Manager)
https://my-mac.your-tailnet.ts.net (for QwenPaw)
Your connection will be fully encrypted, and the browser will show a valid HTTPS certificate (padlock icon). No port forwarding, no firewall changes, and no certificate warnings.
Verify the configuration:
tailscale serve status
Expected output:
https://my-mac.your-tailnet.ts.net
|-- / proxy http://localhost:8080
|-- / proxy http://localhost:8088
Note: If you need to serve on a different port (e.g., 8443), use
--https=8443instead. The MagicDNS URL will include the port number.
7.9 Optional — Public Access with Tailscale Funnel
If you need to share your services with someone outside your tailnet (e.g., a client reviewing a prototype), Tailscale Funnel exposes a local port to the public internet over a stable HTTPS URL — without port forwarding, DNS records, or a public IP address.
Enabling Funnel:
- Go to the DNS page of the admin console.
- Ensure MagicDNS and HTTPS Certificates are enabled (both are required for Funnel).
- Under the Funnel section, turn on Funnel. You need to be an Owner, Admin, or Network admin.
Exposing a service publicly:
tailscale funnel 8080
Expected output:
Available on the internet:
https://my-mac.your-tailnet.ts.net/
|-- proxy http://127.0.0.1:8080
Warning: Funnel exposes your service to the public internet. Anyone with the URL can access it. Do not use Funnel for services containing sensitive data. Always enable password authentication for any service exposed via Funnel.
To stop public access:
tailscale funnel --terminate-on 8080
7.10 Verify Tailscale Survives Reboot
Tailscale runs as a system extension and automatically reconnects on reboot. After restarting your Mac, verify from your client device:
ping my-mac
If the ping succeeds, Tailscale and MagicDNS are active and your server is reachable by name.
Troubleshooting Quick Reference
| Symptom | Fix |
|---|---|
systemsetup: Turning Remote Login on or off requires Full Disk Access privileges |
Grant Terminal Full Disk Access in System Settings → Privacy & Security → Full Disk Access, then quit and reopen Terminal |
pmset -g shows SleepDisabled 1 instead of disablesleep 1 |
Normal on macOS Ventura+. Same setting, different label |
sleep 0 (sleep prevented by powerd, bluetoothd) |
Informational, not an error. Sleep is disabled as intended |
brew: command not found |
Run eval "$(/opt/homebrew/bin/brew shellenv)" or add it to ~/.zprofile as shown in §4.1 |
| Services don't start after reboot | Verify auto-login is enabled; check FileVault is OFF |
mlx-manager not found |
Run eval "$(/opt/homebrew/bin/brew shellenv)" and retry |
| QwenPaw can't reach MLX | Confirm MLX Manager is running: curl http://localhost:10242/v1/models |
| Out of memory errors | Use smaller quantized models (4-bit); reduce context window in MLX Manager |
| Port conflicts | Change MLX Manager port with MLX_MANAGER_DEFAULT_PORT_START env var |
| MagicDNS name doesn't resolve | Verify MagicDNS is enabled in the Tailscale admin console under DNS settings; check Tailscale is running on both devices |
ts.net URL doesn't resolve |
MagicDNS may be off for that resolver; enable it, or use curl --resolve :443: |
| SSH over Tailscale times out | Verify Remote Login is enabled (sudo systemsetup -getremotelogin); check Tailscale SSH policy in admin console |
| HTTPS URL gives certificate error | Verify HTTPS Certificates are enabled in the Tailscale admin console under DNS settings |
| Tailscale Serve not working | Run tailscale serve status to check active configurations; ensure the target services are running locally |
| Funnel URL not accessible | Verify Funnel is enabled in the admin console; check your ACL policy allows Funnel; ensure HTTPS certificates are enabled |





