DVAI Hub

DVAI Hub is the household-utility flavor of distributed inference. Run it on the strongest machine in your house — a desktop, a laptop when it's plugged in, the family Mac mini. Any number of dvai-bridge-powered mobile apps pair with it and offload heavy inference onto it.

You don't have to use it. Mobile apps still work standalone. But when you do:

• Phone-class models that crawl on your phone, burn battery, or evict mid-generation suddenly become snappy.
• A single household Hub serves multiple unrelated apps from multiple family members. Pairings, caches, and audit logs stay isolated per-app.
• External engines already running on the box (Ollama, LM Studio, vLLM, llama-server, llamafile) become available to those apps. The apps don't re-download a model the engine already has cached.

Architecture deep-dive →Developer-fork guide →Migration: v3.0 → v3.1 →

---

Install

macOS

brew install deepvoiceai/dvai-hub/dvai-hub

…or grab the universal .dmg from GitHub Releases.

Windows

winget install DeepVoiceAI.DVAIHub

…or grab the .msi from GitHub Releases.

Linux

Grab .AppImage, .deb, or .rpm from GitHub Releases.

# AppImage
chmod +x DVAI-Hub-*.AppImage
./DVAI-Hub-*.AppImage

# Debian / Ubuntu
sudo apt install ./dvai-hub_*_amd64.deb

# RPM
sudo dnf install ./dvai-hub-*.x86_64.rpm

---

First run

On first launch, the Hub:

1. Starts the embedded HTTP server on a free port. Defaults to the 38883 range; falls back if busy.
2. Advertises itself on your LAN via mDNS as a peer (_dvai-bridge._tcp.local).
3. Lives in your system tray. Click the tray icon — the dashboard opens. Close the window — Hub stays running in the tray.

The dashboard's first-run wizard walks you through:

• Engines — should the Hub surface external engines (Ollama, LM Studio, …)? Default off. You opt in tab by tab from the Engines tab afterwards.
• Auto-start at login — opt in here or later from Settings.

---

Pair your phone

Open your dvai-bridge-powered mobile app on the same Wi-Fi as the Hub. The app's first inference request:

1. Triggers an mDNS handshake to your Hub.
2. Pops an approval modal in the Hub dashboard — "iPhone wants to pair with this Hub on behalf of <your-app>."
3. On approve, a 256-bit pairing key is generated and stored. Every request after that from the same phone is HMAC-signed with that key. No more prompts.

Pairing is per-app. Install a different app — it triggers its own approval prompt and gets its own key. Revoke one app, the others are untouched.

Pairings expire after 30 days of inactivity. Re-handshake is silent if you re-approve.

---

Use it

There's nothing to do. Once paired, your mobile app's inference requests route to the Hub automatically. Subject to the usual offload rules:

• The Hub runs the request locally if it has the model cached.
• If the requested model isn't an exact match, the Hub may substitute a same-shape model with a different quantization — only when you've approved better-quant substitution per-app in Settings.
• If no compatible model is cached, the Hub returns a structured no_capable_device error. Your mobile app's offload policy decides whether to fall back to local inference.

The mobile app's chat client (LangChain, OpenAI SDK, etc.) sees a normal SSE-streamed response — same shape as if it had run locally. The fact that the work happened on your laptop is invisible at the wire.

---

How it works

┌────────────────────┐                      ┌─────────────────┐
│  iPhone running    │ ── mDNS handshake ─→ │   DVAI Hub      │
│  Acme Chat App     │                      │   (laptop)      │
│                    │   /v1/chat/...       │                 │
│  dvai-bridge       │ ── HMAC-signed ────→ │  • peer-mode    │
│  in target=offload │      request         │  • multi-tenant │
│      mode          │ ←── SSE stream ────  │  • engine bridge│
└────────────────────┘                      └─────────────────┘
                                                ↓ (if external
                                                   engine has it)
                                            ┌─────────────────┐
                                            │   Ollama / LM   │
                                            │   Studio / vLLM │
                                            └─────────────────┘

The Hub is built on the same v3.0 distributed-inference primitives that power every other dvai-bridge target — capability assessment, LAN mDNS discovery, HMAC-signed pairing handshake, OpenAI-compat proxy. What's new in v3.1:

• Multi-tenant pairing isolation. A single Hub serves many unrelated apps. Each gets its own pairing key, capability cache, and audit log. The Flavor 2 (developer-fork) build locks to a single appId.
• Strict substitution policy. Request asks for gemma-4-E2B-q4-instruct. Hub has gemma-4-E2B-q8-instruct cached. The policy chooses — refuse (default), or substitute with an explicit warning (per-pairing opt-in). No silent family / version / size / type mismatches.
• External engine bridge. Opt-in framework that surfaces Ollama, LM Studio, vLLM, llama-server, and llamafile as additional backend pools. Each engine's cached models flow through the same canonical- name parser, so the substitution policy can reason about them.

---

Settings

Setting	Default	Where
Port	38883	Settings tab
mDNS service name	dvai-hub	Settings (advanced)
Rendezvous URL	unset (LAN-only)	Settings
Model cache location	OS-specific cache dir	Settings
Auto-start at login	off	Settings
External engines master switch	off	Engines tab
Per-engine toggle	off	Engines tab
Per-app substitution policy	strict	Paired Apps tab

---

Privacy

• Hub-served requests never leave your network. Unless you configure a rendezvousUrl for the internet path.
• Audit logs live on disk in your user-data directory. Nothing is transmitted to DeepVoiceAI.
• The Hub doesn't auto-update by default. Updates come through whatever channel you installed from — Homebrew, winget, direct download.

---

Troubleshooting

My phone can't find the Hub.

• Same Wi-Fi network? Same subnet?
• Corporate / school networks often block mDNS. Try a phone hotspot or your home router.
• Open the dashboard's Status tab — does it show a baseUrl? Visit http://<hub-baseurl>/v1/dvai/health from your phone's browser to confirm reachability.

The pairing modal never appears.

• Look for a tray notification. If notifications are disabled in your OS, the Hub falls back to surfacing the modal on the next dashboard open.
• Is the Hub paused? Check the tray icon. "Resume peer-mode" wakes the embedded HTTP server.

An inference request returns no_capable_device.

• Check the Models tab — is the requested model cached?
• Check the Engines tab — is the engine that has the model enabled?
• Check the per-app Substitution Policy in Paired Apps — is the app allowed to use a different quant?

Full distributed-inference troubleshooting →

---

See also

• Developer fork guide — branded companion for app developers.
• Migration v3.0 → v3.1 — what's new since v3.0.
• Distributed inference — the substrate Hub builds on.
• Self-hosting rendezvous — for the internet path.