Changelog

open a sleeping box and you can start right away instead of watching a progress bar. Your agent answers from what it remembers about you and your work, and says what it will do the moment the machine is ready. — build Amber Kestrel

whatever you discussed while it was starting is waiting for you as your current chat, and your agent picks up where you left off. Say "go ahead" and it knows what you mean. — build Lucky Falcon

the paper on how Avriz picks the right model for every request (avriz.io/eng/paper) was audited against the running code and corrected. Revision 3 had "fixed" the model ladder into price order, which was wrong — it is ordered by capability, and the two disagree; that is now retracted in the errata and right everywhere. The report also documents what routing actually does today: the per-request latency and floor-rung rails, how your plan decides which models are in play, and exactly what a routing decision stores (no message text — and the one content-derived hash it does keep is now named rather than glossed over). It ships in two formats: a typeset PDF and the LaTeX source it was built from. — site update

for people who use assistive technology. The agent’s replies are announced as they arrive instead of landing silently; hands-free mode’s transcript and status now reach screen readers and braille displays, not just speakers; every button has a proper name; the settings toggles, day pickers and menus are fully keyboard-operable, and Escape closes a sheet. Signing up and logging in work with a screen reader too. Pinch-zoom is no longer blocked, and the demos on our home page can be paused. Two gaps we know of and have not closed: the terminal draws to a canvas, so a screen reader is not told when a long build finishes, and we have not yet done a full manual pass with VoiceOver and NVDA. If you use assistive tech, we’d genuinely like to hear from you — we’re working toward WCAG 2.2 AA and we publish the gaps we know about. Watch the 45-second rundown (captioned, with a transcript). — build Brave Meerkat

a brand-new box used to keep finishing its setup for several minutes after the page said it was ready, and anything you typed in that window came back as “the agent is starting up, send your message again”. The agent now starts as soon as it can and the optional extras finish quietly behind it, so a new box is usually answering in around three minutes rather than ten. If you do get in early, Avriz holds your message and sends it the moment the agent is up instead of asking you to retype it. Wandering off during setup is safe too — a box that is still building no longer gets put to sleep half-finished. — build Keen Tapir

if you tap Send while your phone is still finding the network, Avriz now retries for a few seconds and, failing that, keeps the message queued and sends it the moment you’re back, instead of showing “network error” and dropping it. If the message did reach your box before the connection wobbled, it reconnects to that task rather than running it twice. — build Amber Yak

if your phone drops off wifi while the agent is working, the app now says it’s reconnecting instead of sitting frozen, and the moment it gets back it fills in everything you missed. The work never stopped on your box; now you can watch it again as soon as you’re back on signal, instead of waiting for the “task finished” notification to find out what happened. — build Electric Narwhal

Settings → Agent & models now has an engine picker: Goose (the default) or OpenCode (experimental). Switching takes about a minute, keeps all your files, and starts a fresh chat — and you can switch back anytime. Same voice, same tools, same managed model routing either way. — build Gentle Heron

when a request needs the heaviest reasoning (architecture, gnarly debugging), Avriz now routes it to Opus 5 instead of Opus 4.8. Same price per token, so nothing changes on your bill — you just get the newer model on the work that needs it. Pinning a model in Settings still works as before. — build Fuzzy Meerkat

the box quietly learns your preferences as you work, and over time those auto-learned notes pile up as overlapping fragments. Now, once enough accumulate, they get distilled into fewer, sharper facts (duplicates merged, contradictions resolved toward the most recent). Anything you saved to memory yourself is never touched, and a bad distill simply doesn’t apply — memory can shrink, never break. — build Witty Falcon

every task in the Scheduled list now has a pause button. Pausing (with a confirmation so you don't fat-finger it) stops the task from running until you resume it; the row shows “paused” and resume tells you when the next run lands. “Run now” still works while paused for a manual check. — build Rapid Meerkat

waking a sleeping box takes about twenty seconds, but the progress bar was paced against four minutes, so it barely moved and looked stuck. It now matches reality, and if something ever does run long it keeps waiting and lets you in the moment your box is up, instead of quietly giving up and leaving you on a screen that never changes. — build Keen Tapir

Settings → Notifications now has a Send a test button that fires a notification right then, so you can confirm it lands on your device. If it doesn't, it tells you exactly what to fix (not enabled, blocked, and so on) instead of failing silently. Pairs with the pocket-the-phone "task finished" alerts, which now reliably reach you even with the app closed. — build Mellow Gecko

the paper on how Avriz picks the right model for every request (avriz.io/eng/paper) got a corrected revision: the model ladder now matches the price hierarchy throughout, figure captions and equations are fixed, and the downloadable PDF is the new 5-page v3 build with an errata section listing every change. — site update

start something, then lock your phone or switch apps: the task keeps running on your box, the reply is waiting for you when you come back, and you get a notification the moment it finishes. No more losing a long run to a locked screen. — build Rapid Falcon

if a longer task finishes while you're in another app, Avriz sends a notification with the result, so you can pocket your phone and let it work. (Uses the same notification permission as scheduled tasks — enable it once in Settings.) — build Gentle Bison

open your box on your phone and use Add to Home Screen: you now get a real full-screen app with the Avriz icon, not a browser shortcut. — build Gentle Bison

tap the + button next to the composer and add an image (snap a photo of a whiteboard sketch, take a screenshot, or pick a design), a PDF, or a markdown file. Avriz reads it and drops a ready-to-build spec into your message box. It's smart about what it's looking at: a polished mockup gets reproduced faithfully, a flowchart or boxes-and-arrows sketch gets interpreted — Avriz works out the actual app and screens it implies (turning boxes into pages and arrows into navigation) — and a written spec (PDF or markdown) becomes the UI it describes. Review it, tweak if you like, then Send — your agent builds it from there. Big photos are optimized right on your device before upload, so it stays fast even from your phone.

the Stop button now stays put for as long as a task is working, even if you reload the page or switch away and back, so you can always halt a long or runaway task. — build Rapid Falcon

send a message with no signal and it queues with a "sends when you're back online" note, then goes automatically the moment you reconnect. Made for the commute. — build Brave Osprey

in a git project, every reply now has a small undo this turn's changes action: one tap restores your files exactly as they were before that turn (your .env files and ignored folders are never touched). Experiment fearlessly. — build Gentle Bison

a brand-new box now greets you with ready-to-tap ideas, and waking a sleeping box shows exactly what's happening (powering on… starting voice·ide…) instead of a bare spinner. — build Gentle Bison

while the agent reasons, a dim ✦ thinking line shows its train of thought live; tap to read the full reasoning, even after the reply lands. — build Gentle Bison

the engine behind Avriz (Goose) moved up to the latest version, bringing more reliable tool handling and a batch of under-the-hood fixes. It rolls out to your box automatically on its next wake — nothing to do on your end.

the whole interface got a custom icon set drawn to match Avriz's phosphor-green-on-black theme. The old emoji dotted around the app — the tabs, the More menu, the file browser (folders and every file type), and the buttons for commit, upload, new chat, add-a-photo and the rest — are now crisp, glowing line icons that all belong to one family. Purely a visual refresh: everything works exactly as before, it just looks more like one polished tool. — build Brave Otter

every deployed app now has a share button: on your phone it opens the normal share sheet, so the thing you just built by talking is one tap from a friend's screen. — build Brave Osprey

when you add a screenshot, mockup, PDF or spec, Avriz's generated build prompt used to be crammed into the little message box. Now it opens in a full-screen editor: the whole spec laid out in a comfortable, easy-to-read view where you can tweak wording, trim, or add detail, then tap Send to build. Tap back to keep it as a draft in the composer instead. Typing a long prompt yourself? A ⤢ expand button appears so you can pop it open in the same roomy editor. — build Jolly Seal

after the agent replies, a couple of suggested next things to say now appear just above the text box — like “run the tests” or “commit and open a PR.” Tap one to drop it into the composer instead of typing, then send. — build Amber Narwhal

when the agent is reading an answer aloud, the mic button turns into a red ⏹ — tap it to stop talking instantly. Sending a new message stops it too. (Hands-free mode's spoken “stop” / “Hey Goose” still work as before.) — build Sunny Bison

the Usage button now changes color as you use your monthly model allowance (green → amber at 80% → red when it's spent) and shows the percentage once you pass halfway. No surprises at the end of the month. — build Brave Osprey

flip on the Changed filter in the file browser and it now shows a single flat list of every changed file across your repo, each with its folder path — instead of making you dig through nested folders to find them. Tap one to open it, or use the ⋮ menu to diff, stage, or discard.

the file browser now has a ☑ Select button: tap it to pick a bunch of files and folders (or Select all in the current folder), then Move them into another folder or Delete them together — instead of one at a time. Move opens a simple folder picker; delete asks for confirmation first.

tapping a link in the conversation now pops up a small sheet showing the full web address with an Open button (and a Copy link), instead of jumping straight to a new tab — so a mis-tap can't yank you out of your chat, and you can see exactly where a link goes before you follow it.

you can now start with a real Medium (4 GB) dev box completely free: 25 active compute hours a month, Gemini Flash included with a starter token bundle, bring your own voice key if you want speech. There's no bill to run up — when the month's hours are used, the box simply rests until the 1st and your files stay put. Pick "Start free" on the build-your-box page. Seats are limited while we scale it up. One housekeeping rule: a free box left untouched for two weeks gets recycled — we email you 3 days before anything happens (just open the box to keep it), and your account can start a fresh free box anytime after.

hands-free mode no longer needs any tapping: it listens continuously, shows a live transcript as you speak, sends automatically when you pause, reads the reply aloud, then listens again. Steer it by voice — say “stop”, “new chat”, “never mind”, or “exit hands-free”. Speak while the agent is still working and your message queues up. The screen stays awake in hands-free, and tapping the big button interrupts a reply. A glowing 🎤 button in the top bar drops you straight in, on phone and desktop alike. (On browsers without speech recognition, tap-to-talk still works exactly as before.)

when it works out a genuinely reusable, multi-step procedure, your agent can now save it as a skill and apply it automatically in future sessions — so you don't have to explain the same how-to twice. Skills stick around even if your box is rebuilt. See what it has picked up, and switch any of them off, under Settings → Skills. — build Rapid Wombat

hands-free grew a wake-word mode for when you're not alone: say “go to sleep” and it ignores all conversation until you address it with “Hey Goose” — named for the open-source Goose engine that runs your agent, and chosen because it's a real word your phone recognizes reliably. A soft chime confirms it's listening; “Hey Goose, run the tests” works in one breath, and if it ignores you it shows what it heard. Say “open mic” to go back to sending everything. And you can barge in on a spoken reply — “stop” or a new “Hey Goose…” request cuts it off mid-sentence. Your choice of mode is remembered.

real-time voice now uses a conversational model with built-in turn detection, so it replies when you're actually done rather than guessing on a pause. Tune how eager it is — Snappy, Balanced, or Patient — under Settings → Speech recognition. — builds Electric Yak & Witty Quokka

speech recognizers mishear made-up names ("Fugu" becomes "fugue"), so transcription is now tuned to your own workspace. Your box can transcribe with Deepgram's cloud speech engine, biased toward your project's file, folder, repo and brand names, with an automatic fallback to your device's built-in recognizer when it's unavailable or you're offline. On top of that, transcripts are corrected against a vocabulary built from your workspace: "goose hints" snaps to goosehints, "app dot py" becomes app.py. Add your own jargon under Settings → Voice vocabulary; corrections are deliberately cautious, so ordinary words are never rewritten. Pick your engine under Settings → Speech recognition: Auto, Cloud (accurate), or On-device (private). — build Golden Quokka

hands-free is now voice-first: a live waveform moves as you speak, there's no text to read, and your agent tells you out loud when it's working — checking back in every few seconds on longer tasks — so you can keep your eyes off the screen. — build Golden Heron

three fixes for choppy audio: your agent now starts speaking the first sentence while it's still writing the rest; replies are spoken in small sentence-sized pieces that are each fully prepared before playing (and the next piece downloads while the current one plays) — so audio physically can't stutter mid-sentence; and we fixed a bug where transcribing or synthesizing speech briefly froze everything else on your box (the real source of the stutter). On phones the microphone now fully lets go while a reply is playing (holding it open is what made phones garble the audio) — tap the big button to interrupt; on desktop you can still just say "stop".

the file browser now plays audio and video inline (with scrubbing), and lets you pinch or scroll to zoom and drag to pan images, instead of downloading them to look. — build Cobalt Ferret

entering hands-free voice now shows a quick splash: eyes on the road, both hands on the wheel, voice only. It always appears when your box auto-switches to hands-free because it senses you're driving, and once per session otherwise — one tap to continue. We also added a short driving-safety section to the Terms of Service: these voice features are a convenience, not a driving aid, and you're responsible for driving safely and lawfully.

links in your agent's replies, in status messages (like "PR opened: …"), and in messages you paste now open with one tap, in a new tab so you never lose your conversation. Bare URLs turn into links automatically; trailing punctuation stays out of the link so "see avriz.io/docs." goes to the right page.

Auto model routing now improves itself with reinforcement learning, so we spelled out exactly what that means in plain language: it learns from the shape of requests (length, has-code, has-error-trace) and how each routing choice worked — never from the text of your prompts, code, or conversations, which aren't stored in or recoverable from the routing system. Pinning a model in Settings opts you out entirely. See the updated Privacy Policy, Terms, and a new FAQ entry.

you can now keep several conversations going instead of one. Start a new chat any time, switch between them from the Chats tab, and rename or delete them — each one keeps its own context so threads don't bleed together. A search box helps you jump to the right one.

it drives a real browser on your box, so it can open a page, confirm it actually renders, read console errors, and take a screenshot. Handy for checking a deploy looks right or reproducing a visual bug — and the screenshots show up in your Files.

flip on notifications in Settings → Notifications and your box pings your desktop or Android device the moment a scheduled task finishes, even with the tab closed. Email still arrives too. (iPhone support is coming next.)

your assistant now introduces itself as Avriz (say it "AV-riz"), not the old "voice·ide" name, and when you ask what it can do it tells you what it can help you build instead of reciting technical house rules.

link Google in Settings → Connections and your agent can work with your Calendar, Drive, Docs, Sheets and Gmail for you: ask "what's on my calendar today?", pull a figure from a spreadsheet, read a doc, or search your inbox. Reading works the moment you connect; anything that changes or sends things (new calendar events, sheet rows, emails) stays off until you flip a clear "let the agent create/modify" switch — so nothing happens on your accounts without your go-ahead.

running a front-end dev server on your box (Vite, Next, whatever)? Open it over the internet, with hot reload, at your box URL with --<port> added — so a server on port 5173 is at https://your-name--5173.t.avriz.io. No config: it can stay on localhost, no --host flags, and the agent can start it and hand you the exact link. Previews are private to you (a one-time sign-in the first time). Great for watching the UI update as you and your agent build; use Deploy for a finished, permanent app URL.

upload a coding standard, spec, or convention (More → Knowledge) and your agent treats it as authoritative and follows it. It also keeps a growing knowledge base of your project as it works, so it gets sharper about your codebase over time.

turn on "Hands-free while driving" (Settings) and when your phone senses you're on the road it switches to a big talk-to-your-agent view with replies read aloud, so you can keep going by voice with your eyes up. Your location never leaves your phone — it's only used on-device to sense movement. (Works while the box is open on your screen.)

if a model provider retires or breaks a model on their side, your agent now reroutes to another model mid-conversation instead of going silent. This fixed today's brief outage within the hour.

the Files tab has a search box that finds any file across your whole workspace by name, however deeply it's nested; clear it to go back to browsing.

fixed a problem where a running agent could stall partway through a task with a "network error" while using its tools. Those steps now run through to completion.

fixed a problem where an image from the browser tool (like a screenshot of your app) could error out the reply or get silently ignored before it reached the model. Now it comes through, so "look at this page and tell me what's off" works.

open a PDF in the Files tab and it now displays right there, all pages, scroll straight through — no download needed. Prefer the file itself? A Download button is one tap away.

in the Files tab, tap ⋮ on any folder and choose "Download as zip" to grab everything in one file. It leaves out bulky, rebuildable folders like node_modules (while keeping your git history), so the download stays small and quick.

on a computer these now sit right in the top toolbar instead of only the phone menu, so you can switch between conversations, open your knowledge base, or set up a scheduled task in one click.

turn on "Show which model answered each reply" (Settings → Model) and each reply gets a small tag showing the model your Auto routing picked, so you can watch it send easy questions to a cheap model and hard ones to a stronger one. Off by default.

if you leave a tab open and your box powers down after sitting idle, you now get a clear "your box is asleep" pop-up with a one-tap Wake button, instead of things quietly not responding.

the usage heads-ups ("80% of your allowance", "allowance reached", "credits running low") used to arrive as a plain line of text; now they carry the glowing a·v·r·i·z wordmark, a usage meter you can read at a glance, and a one-tap button to the right place — matching the welcome, invite and reset emails. Payment-failed and reconnect emails picked up their action buttons too.

the Goose engine is now every box's brain, so waking your box is one tap (no more engine picker) and every improvement we ship lands in the agent you're actually using. It also knows the platform now: it can publish your app with one deploy command, open GitHub PRs for you, and respects the house rules (it won't touch the port your IDE runs on). And if your plan's included usage runs out mid-task, it tells you clearly instead of stopping with a confusing error.

it quietly picks up lasting preferences and project conventions from normal conversation (like "use python3" or "deploy to main") and remembers them across sessions, so you don't have to repeat yourself. Settings → Memory shows what you asked it to remember separately from what it learned on its own, with edit and delete on every fact — and auto-learned facts quietly fade after ~2 months unless they keep coming up, so stale preferences don't linger.

Files and Code are now one tab: browse folders with live git status badges, filter to just what changed, view diffs, stage/unstage/discard, and commit or open a PR — all in one place. One fewer tab on mobile.

spoken replies no longer read code aloud (code stays on screen, the voice sticks to plain language), long conversations now compact themselves automatically so the agent stays fast, and when a task turns out harder than it first sounded, the agent quietly steps up to a stronger model mid-task.

the in-app terminal was failing to connect on any box reached by its memorable name (a security check didn't recognize the friendly URL). It now connects reliably; nothing to do on your end.

the whole app, especially Settings, now works properly on phones.

each request auto-picks the right model and falls back instantly if one is busy.

line up follow-ups while the agent works, and interrupt a task mid-run.

we fixed cases where it could reply with nothing, and sped up responses.

your box now lives at your-name.t.avriz.io (old links still work).

apps you ship to *.avriz.app get a valid certificate right away.

at avriz.io/docs — guides for the box, billing & settings.