Domo
Domo is an on-device AI agent built on Gemma 4 that commands a cloud fleet.
Core trick
A small on-device Gemma model routes tasks to cloud agents and screens every returned result for prompt injection before it enters its context.
Description
Domo is an on-device AI agent built on Gemma 4 that commands a cloud fleet. A small Gemma model runs entirely on your machine as the orchestrator — it listens, reasons, remembers, uses tools, and dispatches background agents (research, code, image generation), escalating to the cloud only when it must. Everything works offline: when the network drops, Domo keeps reasoning, recalls from local memory, tells you plainly that it's degraded, and queues cloud work until you're back.
What makes Domo different is the trust boundary. It sits as a membrane between you and the fleet: every result an agent brings back is screened for prompt-injection before it can touch Domo's context, so a poisoned web page can't hijack the agent — and only minimal, screened context ever leaves your device, so your private data isn't shipped off to cloud vendors. A cloud orchestrator structurally can't offer this; Domo can, because the brain lives on your device.
Domo is role-adaptive — the same on-device brain becomes a developer, a PM, a marketer, or a researcher, reaching for the right skill. Hand it a photo and say "turn this into a launch ad," and it writes the copy and generates the image in seconds.
Built for the real world: across regions with spotty connectivity, high data costs, or sensitive data in health, finance, and personal privacy, sending everything to a server is a non-starter. Domo's offline resilience and privacy aren't features bolted on — they're the architecture.
Powered by on-device Gemma 4, a managed-agent fleet, a two-tier memory, and NB2 Lite for instant creative generation.
README excerpt
# Domo
A local-first agentic OS. A Gemma model on your device listens, remembers, uses tools,
and commands a fleet of background agents for you — escalating to the cloud only when it
needs to. The assistant survives offline; the work resumes when you reconnect.
Built for the Google DeepMind Bangalore Hackathon (Gemma 4 · Gemini Live API · Managed
Agents · Antigravity · Omni Flash).
## Why
Your assistant should run where your context lives: on your device. Private by default,
usable offline, cheap enough for a low-end phone. The cloud is muscle you rent when the
task needs it, not the place your assistant lives.
## Architecture (two-stage tool calling)
Small models are unreliable at *choosing* a tool but excellent at *formatting* a call once
the tool is chosen. So Domo splits the job:
```
request ──► Gemma 4 (reasoner) ── picks the tool, states intent in plain language
│
▼
functiongemma-270m ── given ONE tool, emits a structured, parseable call
│
▼
tool executor ── real side effect + write to local memory ──► reply
```
Measured: functiongemma formats calls with 100% parse success but mis-routes ~40% when
handed many tools. Handing it a single pre-chosen tool removes the routing failure while
keeping the perfect formatting. See `experiments/`.
## Stack
| Layer | Component | Runs |
|---|---|---|
| Reasoner brain | Gemma 4 E4B (4-bit, MLX) | on-device |
| Tool-call head | functiongemma-