Offline AI assistant
The network drops. The assistant doesn't.
Almost every AI assistant is a thin client for someone else's datacentre. Lose the connection — on a train, on bad hotel WiFi, during an outage — and it stops being an assistant at all.
Vyra treats the local model as the floor of the system rather than a novelty. A 5-tier router picks the right model per request, and the bottom tier runs on your own machine via Ollama — so when connectivity disappears, routing collapses to local instead of failing.
The five tiers
Local
An Ollama-hosted model running on your own machine. Always available, no network required — the tier everything else falls back to.
Fast
Small cloud models for high-frequency, low-stakes work where latency matters more than depth.
Creative
Models tuned for drafting and generation, where variance is a feature rather than a defect.
Thinking
Larger reasoning models for multi-step problems that justify the extra time and cost.
Ultra
The heaviest tier, reserved for work where being right matters more than being quick.
Why the routing decision is made per request rather than per session, and what the real quality trade-offs are, is covered in local vs. cloud models.
What still works with no connection
Voice keeps listening
Wake-word detection and speaker identification run locally, so the assistant doesn't stop hearing you the moment the network drops. What changes is which model answers, not whether one does.
Memory stays intact
Episodic memory and the world model live on your machine. Retrieval — both full-text and semantic — works offline, which means the assistant still knows who and what you're talking about.
Local actions still execute
File operations, device control on your own network, and locally-defined automations don't need a round trip to a datacentre and don't stop when one is unavailable.
Cloud work queues rather than fails
Steps that genuinely require an external service are held rather than errored, and resume when connectivity returns — so a dropped connection costs you a delay, not a lost task.
Memory surviving the outage is the part that matters most — it's why the assistant is still useful offline rather than merely responsive. That system is described in full under AI assistant with memory, and the automation side under AI workflow automation.
Common questions
Can an AI assistant really work offline?
Partly, and the honest framing matters. Anything that depends on an external service — sending a message, calling an API, searching the live web — needs a connection by definition. What can run offline is the reasoning layer, via a local model, plus wake-word detection, memory retrieval and any action that stays on your own machine or network. Vyra is built so that losing connectivity degrades capability instead of stopping the assistant.
Which local model does Vyra use?
The local tier runs through Ollama, which means the specific model is a choice rather than a hard-coded dependency — you can run a small, fast model on modest hardware or a larger one if your machine supports it. The router treats it as the always-available floor beneath the cloud tiers.
How does it decide between a local and a cloud model?
A 5-tier router picks per request based on what the request actually needs — latency sensitivity, reasoning depth, and cost — rather than sending everything to the largest available model. Connectivity is one input among several: when the network is gone, the local tier is the only candidate, so routing collapses to it automatically instead of erroring.
Is a local model worse than a cloud model?
For hard multi-step reasoning, generally yes — that's why the higher tiers exist. For the bulk of everyday assistant work (triage, drafting, classification, retrieval over your own memory) the gap is much smaller than the benchmark difference suggests, because those tasks are bounded and the context is already local.
Does running locally mean my data stays on my machine?
When the local tier handles a request, the prompt and the retrieved context don't leave your machine. When a cloud tier handles it, they do — that's inherent to using a hosted model. The useful property of the router is that it makes this a visible, per-request decision rather than an invisible default.
Vyra is in closed alpha, with a Founders Beta ahead of public launch in 2026.