Settings & keys
There are two settings surfaces, and knowing the split saves you hunting: account settings at avriz.io/settings (plan, usage, credits) and box settings inside your box (keys, model, tools, memory). The box also mirrors your account under Settings → Account & box — plan, usage meters and credit packs — so on a phone you never have to leave it.
Account settings — avriz.io/settings
- Plan — your free tier at a glance: box, model routing, voice. Nothing recurring, no card on file.
- This month's usage — allowance %, tokens, voice minutes, compute hours. The free tier includes about two hours on the best models; past that the agent keeps working on the fast ones (Gemini, Haiku) rather than stopping. See Billing for what happens past the included amounts.
- Credits — your balance; buy packs from the dashboard.
Box settings — inside your box
Open your box and hit Settings (under More on a phone). This is where the agent's behavior is configured:
API keys (BYO)
| Key | What it enables |
|---|---|
| LLM key + base URL + model | Run the agent on your own OpenAI-compatible provider (OpenAI, Anthropic, a proxy, a local gateway…). When set, your key is used instead of managed models — no metering. |
| ElevenLabs key | Voice bills to your ElevenLabs account instead of the included/metered voice. |
| GitHub token | Lets the agent clone private repos, push, and open PRs as you. A fine-grained token scoped to the repos you'll use is best. |
| AWS credentials | Optional — give the agent the ability to manage your AWS resources (access key, secret, region). |
Agent model picker
On managed plans, choose Auto (a judge grades each request and routes it up the Gemini → Fugu → Sonnet 5 → Opus 5 → Fugu Ultra ladder) or pin one model for everything. Billing follows the model actually used — details in Billing & credits.
Agent engine
Settings → Agent & models also picks the engine itself: Goose (recommended, the default) or OpenCode. Switching takes about a minute and starts a fresh chat; your files, memory and keys are untouched.
Notifications
Settings → Notifications enables push on the current device — you get a notification when a scheduled task finishes, even if the app is closed. Works on desktop and Android; each device opts in separately, and there's a "send a test" button to confirm it.
Memory
The Memory card shows the durable facts the agent recalls every session, in two groups: Saved by you (things you told it to remember) and Picked up automatically (preferences it learned from normal conversation). Edit or delete any fact. Auto-learned facts fade out after about two months unless they keep coming up — editing one keeps it for good. Memory survives sessions, box sleeps, and even a box recycle — it's stored off-box, scoped to your account.
Speech recognition
Choose how your speech is turned into text: Auto (accurate cloud recognition when your box has it, your device's built-in recognizer otherwise — with automatic fallback), Cloud (accurate), or On-device (private), which keeps your audio on your device. The cloud engine is tuned to your project's vocabulary so it gets file, folder, repo and made-up names right. For hands-free, an end-of-turn sensitivity control — Snappy · Balanced · Patient — sets how soon it decides you've finished talking and sends; nudge it toward Patient if it ever cuts you off, or Snappy if it feels slow to reply.
Feature toggles
Switch parts of the box on or off: voice, web search, terminal, git, and auto-run (whether the agent may execute commands without asking first).
Agent tools
Your agent is powered by Goose. Manage its tools, MCP servers, recipes and skills from the box's Settings tab. See Inside your box.
Skills — your agent learns your workflows
The Skills card lists the reusable procedures your agent can apply automatically. Alongside a built-in starter set, your agent writes its own: when it works out a repeatable, multi-step how-to, it saves it as a skill and reuses it in future sessions — so you don't re-explain the same task. Skills survive sessions, box sleeps, and even a full box recycle. Each shows whether it was self-authored or saved by you, and you can disable or delete any of them. (Skills that could expose a secret or run a destructive command are refused automatically.)