Inside your box

Your box at https://<name>.t.avriz.io is the actual IDE: agent, editor, terminal, files, and deploys. The tabs are Agent · Files · Term · Apps · Settings — along the top on desktop, a bottom bar on a phone. Chats live in the toolbar (the Chat ▾ switcher next to New), and Help opens these docs.

The Agent tab

• Voice — tap Talk, say what you want, and the agent replies out loud (and in text). Everything works typed, too. Transcription knows your project's vocabulary: file, folder and repo names from your workspace, so "open app dot py" becomes app.py and made-up names stop being misheard. Add your own words under Settings → Voice vocabulary. Choose how speech is recognized under Settings → Speech recognition: Cloud (accurate) uses a cloud engine tuned to your vocabulary, On-device (private) keeps audio on your device, or Auto picks the best available and falls back automatically.
• Hands-free mode — enter it from Settings (or let it switch on automatically while driving) and go fully no-touch and eyes-free: the screen is just a live waveform that moves as you speak — no text to read — and the agent talks to you, telling you out loud when it's working and checking back in every few seconds on longer tasks. It listens continuously, knows when you've finished talking and sends on its own (tune how eager under Settings → Speech recognition: Snappy · Balanced · Patient), then reads the reply aloud before listening again. Say "stop", "new chat", "never mind", or "exit hands-free" to steer it by voice. With passengers around, say "go to sleep" to enter "Hey Goose" mode — it then ignores everything not addressed to it, wakes on "Hey Goose" (a soft chime confirms; your agent's engine is the open-source Goose, hence the name), and "open mic" switches back. You can interrupt a spoken reply mid-sentence with "stop" or a new "Hey Goose…" request. The screen stays awake throughout.
• Real tools — the agent edits files in your workspace, runs shell commands, uses git, and can schedule jobs with cron. You watch each step appear above the final answer.
• Scheduled tasks tell you when they're done — ask for something on a schedule ("run the tests every morning") and you'll get an email with the result when it finishes, plus a push notification on any device where you've enabled them (Settings → Notifications) — even with the app closed.
• Sees your app — the agent can drive a real browser on your box: open a page, confirm it actually renders, read console errors, and take a screenshot (screenshots show up in your Files). Handy for checking a deploy looks right or reproducing a visual bug.
• Markdown replies — answers render with formatting, code blocks and links.
• Continuity — the conversation resumes where you left off, even after the box sleeps or you switch devices.

Memory

The agent has durable memory: tell it “remember that we deploy Fridays only” and it stores the fact — surviving new sessions, box sleeps, even a full box recycle. It recalls facts automatically each turn, and quietly picks up lasting preferences from normal conversation. Manage them under Settings → Memory in the box (saved-by-you and picked-up-automatically, each editable and deletable), or just ask the agent to forget something. See Settings for details.

Which model answers?

With Auto (the default), a fast judge grades each request and routes it up a ladder — Gemini Flash for chat, Sonnet 5 for real coding, Opus 5 / Fugu Ultra for the hard stuff. Pin a specific model any time in the box's Settings (model picker: Auto / Haiku / Sonnet / Opus / Fugu / Gemini…). If you've set your own API key in Settings, it is always used instead — your key, your provider, no metering.

Auto keeps getting smarter: we continuously measure how well each routing choice worked — did the model handle it in one go, or did the request need bumping up to a stronger one — and use that to tune the picks over time. The goal is simple: your allowance stretches further when easy work stays cheap, and hard work reaches a capable model on the first try.

Chats

You can keep several conversations going at once instead of one long thread. The Chat ▾ switcher in the toolbar lists them all (with New right beside it):

• Start a new chat any time, and switch between them freely — each chat keeps its own context, so unrelated threads (say a bug fix and a new feature) don't bleed into one another.
• Rename or delete any chat; new ones are auto-titled from their first exchange so the list stays scannable.
• Search your chats to jump straight to the one you mean.

Terminal

The Term tab is a full shell on your box (xterm) — the same machine and user the agent works as. git, node, python3, npm, build tools and cron are preinstalled; install anything else with the package manager. What you change in the terminal, the agent sees, and vice versa.

Files — a git-aware file manager

The Files tab is a folder-tree browser with git built in as a first-class layer — it replaced the separate Code tab, so browsing files and working with git now live in one place.

• Browse & edit — folder tree with breadcrumbs; create, rename, move and delete; edit any text file in place. Markdown gets a Preview toggle; images preview with zoom (pinch or scroll to zoom, drag to pan); audio and video play inline with scrubbing; and PDFs preview right in the app — all pages, scroll straight through, no download needed (a Download button is still one tap away).
• Bulk actions — tap the ☑ Select button to pick many files and folders at once (or Select all in the current folder), then Move them into another folder or Delete them together in one go — no more one-at-a-time. Move opens a folder picker; delete asks first.
• Git status at a glance — every entry shows its state (modified, added, deleted, renamed, untracked) and folders roll up “changes inside.” Flip on Changed and the browser switches to a single flat list of every changed file across the repo, each with its folder path — so you see exactly what you've touched without drilling through nested folders. Tap one to open it, or use the ⋮ menu to diff / stage / discard.
• Diff, stage, commit, PR — view a file's diff, stage / unstage / discard, then Commit or Open PR. The diff comes with a plain-language AI summary and a suggested conventional-commit message you can accept in one tap.
• Search — the search box finds any file across your whole workspace by name, however deeply it's nested; clear it to go back to browsing.
• Upload from your device and download anything back out.

Knowledge — give the agent your standards

Memory (above) is for short, personal preferences. Knowledge is the other half: durable, project-level ground truth that lives with your repo.

• A living project knowledge base — as the agent works, it maintains a set of markdown notes about your codebase (schemas, conventions, API contracts, deploy runbooks), so it gets sharper about your project over time. It's plain files, versioned with your repo and human-editable (Google's Open Knowledge Format).
• Bring your own standards — open More → Knowledge and upload a coding standard, spec, or style guide. The agent treats an uploaded document as authoritative and follows it — not a suggestion it might weigh, but a convention it applies.

The Goose engine

Your box's brain is the open-source Goose coding agent, running with managed model routing, durable memory, and voice built in — plus its own ecosystem. (Prefer a different engine? Settings → Agent & models can switch the box to OpenCode; switching takes about a minute and starts a fresh chat. Goose is the recommended default.)

• Extensions / MCP servers — enable or add tool servers from a structured UI (no YAML). The browser (Playwright), GitHub, docs lookup and your knowledge base all ride on this.
• Recipes — save reusable task playbooks; edit them right in the page.
• Skills — reusable how-tos the agent applies automatically when they fit. It ships with a starter set and writes its own: when it works out a repeatable procedure, it saves it as a skill and reuses it in later sessions — and skills persist even if your box is rebuilt. See what it has learned, and switch any off, under Settings → Skills.

GitHub & deploying apps

• GitHub — add a GitHub token in the box Settings and the agent can clone, commit, push and open PRs as you.
• Deploy to the web — ask the agent to deploy a project and it publishes to https://<app>.avriz.app. Builds use your Dockerfile if present, otherwise auto-detect (nixpacks), otherwise serve static files. Each app runs in its own container with its own URL; redeploy any time, logs and restart included.

Preview a dev server (live front-end)

Running a front-end dev server on your box? You can open it in your browser over the internet — with hot reload — at a preview URL built from your box address:

• The URL — take your box URL <you>.t.avriz.io and add --<port> to the first part. So a dev server on port 8990 is at https://<you>--8990.t.avriz.io.
• Zero config — start your dev server however you normally would (e.g. npm run dev). It can stay bound to localhost; the box rewrites the Host for you, so you don't need --host or an allowedHosts tweak. Hot-module reload works over the tunnel.
• Only you can see it — a preview is private to your account. The first time you open one you'll sign in to avriz.io once to prove it's your box; after that it just loads.
• Any dev port — any port from 1024 up works, except the box's own services (the app, Postgres, Redis and the agent engine are off-limits).
• The box has to be awake and your dev server has to be running. If the page 502s with “dev server not reachable,” start (or restart) the server on that port.
• This is for development preview. To publish a real app with its own permanent URL, use Deploy to the web above (*.avriz.app).

HMR tip: most setups work as-is. If your framework hard-codes its hot-reload socket, point it at the preview host (for Vite, server.hmr.clientPort = 443).

Help & version

The Help button in the toolbar opens these docs in a new tab. Under More, the box shows the release it's running — a memorable codename (e.g. Crimson Heron) with the exact build id beneath it — so you always know which version you're on. New releases roll out automatically; a sleeping box picks up the latest the moment it wakes.

Sleep & wake

Idle boxes power down automatically — that's why the free tier is workable: you don't pay to idle. The first request wakes it: you'll briefly see the wake screen, then everything is exactly as you left it — files, repos, running config, and your chats. Typical wake is 30–60 seconds. Your 25 included box hours a month only count while the box is actually running; at the cap the box rests until the month resets, or wakes as soon as you add credits.