Router that virtually distributes inference across connected devices in the home.
NVIDIA Personal AI Router — macOS install
This tarball is the portable backend bundle. The shipped distribution places
its graphical UI alongside this bundle in the same installation directory. The
UI launches nvpair-ui-broker, which drives the API and supervises the workers
under bin/. See Unsigned builds if macOS blocks the
binaries on first launch.
1. Extract
tar xf NVIDIA-Personal-AI-Router-<version>-darwin-arm64.tar.gz
cd NVIDIA-Personal-AI-Router-<version>
You should end up with a top-level directory containing bin/ (the worker
binaries) and this INSTALL.md.
2. Run
The bundled UI normally launches the broker from the same installation
directory. For backend-only development or direct API access, run it yourself.
It supervises the other workers (which it expects as siblings in the same
bin/ directory) and speaks newline-delimited JSON-RPC 2.0 over stdio (or a
Unix socket via --ipc):
./bin/nvpair-ui-broker
The bundled UI connects to the broker over this contract. Other clients can use
the same API; see the nvpair-ui-broker README for its JSON-RPC surface.
3. Uninstall
rm -rf /path/to/NVIDIA-Personal-AI-Router-<version>
The workers store configuration in
~/Library/Application Support/Nvidia Corporation/Personal AI Router/ (manual-node list,
log-level preference, cluster identity/pins, etc.). Remove that directory too if
you want a fully clean slate:
rm -rf "$HOME/Library/Application Support/Nvidia Corporation/Personal AI Router"
Downloaded models are not in either directory. Each engine keeps them in its own
store — ~/.ollama, ~/.lmstudio/models, and ~/.llamacpp — so deleting the
bundle or the configuration directory leaves them in place.
Requirements
- macOS 12 (Monterey) or newer.
- 64-bit Intel (
amd64) or Apple Silicon (arm64). The tarball is built per arch; pick the one that matches your Mac.uname -mon the Mac tells you which:x86_64→ amd64,arm64→ arm64. - A working mDNS responder on UDP 5353. macOS's built-in
mDNSResponderis fine; the discovery library coexists with it without configuration.
Unsigned builds
A tarball built with installer_build.sh contains unsigned binaries.
Gatekeeper may block unsigned downloads on first launch. If it does, clear the
quarantine attribute once:
xattr -dr com.apple.quarantine /path/to/NVIDIA-Personal-AI-Router-<version>
Troubleshooting
- No nodes discovered: confirm UDP 5353 isn't blocked by System Settings →
Network → Firewall, and that other machines on the LAN are actually
advertising. Run the broker with
--log-level debugto see live discovery logs on stderr. - "Bind: address already in use": the proxies listen on
11434(Ollama),1234(LM Studio), and8080(llama.cpp); another instance of the proxy or another process is holding one of them. Find the offender withlsof -nP -iTCP:11434 -iTCP:1234 -iTCP:8080 -sTCP:LISTENand stop it.
Report issues at the project's tracker. Include the broker's --log-level debug
output and your macOS version (sw_vers) and architecture (uname -m).