Your Sona releases.

1.5.0

Faster, steadier, better desktop

1.0.0

Sona 1.0 (private)

0.9.5

Release candidate

0.9.2

Desktop updates

5.0.0 — Release candidate — accounts, installer, loading screen, plugins

• Sona 5.0: the release candidate — everything since 2.0 in one release, tested on the live private server
• Accounts: create an account by email or with Google, Discord or GitHub (invite-only), confirm your email, reset a forgotten password, and connect several login methods to one account — Settings → Account
• A new loading screen on the web, the installed app and the desktop app that shows what Sona is really doing (starting, checking your session, signing you in, reconnecting)
• Plugins: the owner installs reviewed plugins from Sona's catalogue (Calculator, Clock & dates, Unit converter to start); use them as /calc, /time, /convert or as agent tools — Settings → Plugins
• Windows desktop installer with the AnxLabs look (welcome, install options, progress, finish, clean uninstall); desktop app 5.0.0
• Release notes and the launch package are ready; Sona stays private until the owner says go

4.9.0 — Launch rehearsal and promotional drafts

• Launch rehearsal: every step of the public-launch runbook checked on the private server — nothing made public
• Draft announcements, a short-video script and a website — kept private until the owner's GO
• deploy/launch-rehearsal.py: a read-only go/no-go check before any launch

4.8.0 — Final acceptance and recovery testing

• Recovery tested on the live server: rolled back to the previous version and forward again (19 s each), restored a backup copy, and ran a load test with no errors
• Fixed: the emergency server rollback script would have failed — found by the drill, now covered by a test
• Recovery record and how to repeat it: docs/operations/recovery-drill-4.8.md

4.7.0 — External beta readiness

• Settings → Help: send feedback from inside the app — it includes the app version, and the open chat only if you tick it
• Admins see and triage beta feedback; the owner sees a live beta-readiness checklist
• Beta checklist for the owner (docs/beta/checklist.md) — no one is invited without your GO

4.6.0 — Documentation, licensing, privacy and support

• Export everything: the full export (Settings → Data) now includes projects, Project Brain facts, learned preferences, feedback, Lab cases, research and team runs, workflows, compute nodes, webhooks and settings — never secrets
• Draft privacy policy, terms of use and support guide (not published), in docs/legal and docs/support.md
• Third-party license inventory for every package Sona ships (THIRD_PARTY_LICENSES.md), with the licensing decisions that remain

4.5.0 — Installers, updates and rollback

• New installer for running your own Sona server — checks the machine first, generates secrets, keeps it private
• The web app can be rolled back to the previous version in seconds (tested on this server)
• Operator guide for installing, updating and rolling back: docs/operations/install-update-rollback.md

4.4.0 — Security and abuse-resistance audit

• Security audit: static analysis, dependency audits and abuse tests — results in docs/security/audit-4.4.md
• Sona now slows down anyone sending far too many requests, so no one account can exhaust the server
• Malicious documents that would expand to gigabytes are refused

4.3.0 — Account security and data isolation

• Your data is walled off from every other account — now proven route by route against another account's real ids
• Settings → Privacy → Delete my account removes the account and everything in it, files and workspaces included
• The command sandbox is never shared with other accounts, whatever the server settings say

4.2.0 — Bring your own compute and providers

• Bring your own compute and providers: every account uses its own GPUs (Settings → Compute) and its own keys
• This server's hardware, keys and subscriptions stay its owner's — other accounts get only what's shared on purpose (web search and voice)
• Settings → Compute shows what's shared with you and what to bring yourself

4.1.0 — Hardware and platform compatibility

• Settings → General → This device: see which Sona features your browser supports (install, offline, voice, screen sharing…)
• Admins also see what the server can run and which optional services are connected
• A new check script tells operators whether a machine can host Sona — and Sona refuses to start, with a clear reason, on a system that's too old

4.0.0 — Onboarding and accessibility

• A short welcome for new accounts: picking models, bringing your own compute and keys, and how your data stays yours
• Keyboard-friendly: a clear focus ring, “Skip to the conversation”, and dialogs keep focus inside and give it back when they close
• Easier to read: secondary text now meets the WCAG AA contrast level on every surface

3.9.5 — Private acceptance of the Learning Era

• Everything new in 3.x re-checked: accounts, learning, Self-Learning Core, Lab, copilots and research stay with your account
• Only you (the owner) can see or apply Sona's improvement proposals and run the Learning Laboratory
• Backups restore cleanly; Sona stays fast under load

3.9.0 — Learning security and reliability

• Learning is guarded: text that tries to instruct Sona about its own rules can't become a preference, a project fact or an improvement
• Project facts reach Sona as facts, not instructions — and everything learned is checked again on its way into a prompt
• Research, team and evaluation runs now have time limits, so nothing can hang

3.8.0 — Local AI research assistant

• Tasks → Research: ask a research question and your own model writes a cited report from web sources
• Sona plans the searches, reads the pages, notes what each source says, then writes the answer with [n] citations
• Citations are checked — invalid numbers and paragraphs without a source are shown; paid models are never used

3.7.0 — VENTARA Creative Copilot

• VENTARA Creative Copilot (Tasks): NPCs, techniques, quests, items, lore and regions for Ventara, as ready-to-use JSON
• Every result is checked against Ventara's design — classless, no levels, no XP, no skill points — and rewritten once if it slips
• Ground it in your VENTARA project's files, then copy the JSON or save it to the project

3.6.0 — Developer Copilot

• Developer Copilot: import a public Git repository into a chat's agent workspace (GitHub, GitLab, Codeberg, Bitbucket)
• Let the agent explain it, write tests or fix failing tests — new developer skills /explain-repo, /write-tests, /review-changes, /fix-tests
• Review every change as a diff and download it as a .patch for your own checkout — Sona never pushes anything

3.5.0 — Local model adaptation research

• Research: could Sona fine-tune a local model on what it learned from you? Findings in docs/research/local-model-adaptation.md
• Settings → Learning → download your corrections and 👍 replies as a dataset file (secrets left out) — for you to keep
• Sona itself never trains or swaps a model; that stays your decision

3.4.0 — Improvement approval and rollback

• Owner only — apply one of Sona's proposals with one click, after evaluating it in the Learning Laboratory if you like
• Every applied change can be rolled back, restoring exactly what it replaced
• Rollback won't silently undo your own later edits — Sona asks first; everything is in the security log

3.3.0 — Learning Laboratory and evaluations

• Owner only — Learning Laboratory (Settings → Learning): eval cases that measure whether Sona gets things right
• Build cases from your corrections in one click, or write your own (must contain…, at most N words)
• Evaluate a proposal before applying it: Sona runs the cases as things are and with the change, and shows the difference

3.2.0 — Owner-only Self-Learning Core

• Owner only — Sona's improvement proposals (Settings → Learning): Sona looks at your ratings, corrections and Arena votes and proposes concrete changes
• A proposal is either a Sona Auto rule (“use this model for code”) or one line for your custom instructions — nothing else
• Each comes with its reason and the evidence behind it; nothing changes on its own

3.1.0 — Learning Core

• 👍 / 👎 on every reply — a 👎 can say what Sona should do differently
• Settings → Learning: Sona suggests preferences from your notes; only the ones you switch on are used
• Edited replies are kept as corrections, and Sona Auto prefers models you rate well for each kind of task
• All yours to see and delete — or turn learning off entirely; nothing trains a model

3.0.0 — Accounts, roles and Supabase Auth

• Accounts and roles, checked by the server: owner, admin and member — every account sees only its own data
• Settings → Accounts (owner/admin): see accounts, make someone admin, disable an account, create invite codes, read the security log
• Ready for Supabase Auth: sign in or create an account with email once a Supabase project is connected — off on this server for now
• New accounts are off by default; invite-only or open sign-up is a server setting, and owner access is never granted by sign-up

2.9.5 — Private acceptance of the 2.x ecosystem

• Everything new since 2.0 re-checked: sync, compute nodes, teams, webhooks, Project Brain and Echo stay with your account
• Data added in 2.x is cleaned up with what it belongs to — chats, projects and accounts
• Backups restore cleanly; Sona stays fast with chats, an arena run and a workflow at once

2.9.0 — Performance

• Sona opens faster: the model list no longer waits for every GPU to answer — it shows what Sona already knows and refreshes in the background
• Updates download about two thirds less: the libraries Sona is built on are now cached separately and kept across updates
• One request fewer on every start

2.8.0 — Echo across devices

• Echo across devices: while Sona Live is on, your other devices show “Sona Live is on in … — Continue here”
• Continue here moves the voice conversation to that device, in the same chat; the first device stops listening and says where it went
• Only presence travels between devices — no audio, transcripts or screen images; every turn stays a normal chat message

2.7.0 — Project Brain and verified knowledge

• Project Brain: each project keeps the facts it must get right — open a project → Project Brain
• Check a fact against the project's files: Sona marks it verified (with the quoted line), disputed, or not in the files
• Propose facts from a chat: Sona suggests decisions and names from it; nothing is used until you accept and confirm it
• Only verified facts are given to Sona in the project's chats, with their source; old ones are flagged for review

2.6.0 — Offline-first web app

• Offline-first (opt-in per device): Settings → Privacy → Keep chats on this device for offline use
• With it on, the installed app opens without a connection — or while the Sona server is down — and shows your chats
• Messages you send while offline wait in an outbox and go out in order when you're back online
• Off by default; turning it off or signing out deletes the copy from the device

2.5.0 — Sona Auto 2.0

• Sona Auto 2.0 sorts every message into a kind of task — code, images, writing, reasoning, quick questions
• Settings → Models → Sona Auto rules: pick the model Auto uses for each kind of task
• Auto learns from you: models you vote up in the Model Arena get picked more, models whose replies keep failing less
• Rules never override safety: paid models only when allowed, secrets stay local — and Auto tells you when it skipped a rule

2.4.0 — Skills and integrations

• Skills: six starter skills to add with one click (summarize, review-code, research, …)
• Share or back up skills as a skill pack file: Export all / Import pack under Assistants → Skills
• Settings → Integrations: webhooks tell other apps when a workflow or team run finishes — signed, so they know it's Sona
• Webhooks send only status, names and times unless you choose to include the generated text

2.3.0 — Multi-agent collaboration

• Tasks → Team: give a goal to a small team of models — a planner splits it, workers do the parts, a reviewer checks them
• Workers on different GPUs work at the same time; the reviewer can send weak parts back once before the final answer
• See every contribution — who wrote what, with which model and how long it took — then continue the result in a chat
• Team members only think and write: no tools, files or web, so a run never needs your approval; free and local models only

2.2.0 — Secure distributed compute

• Settings → Compute: use the GPU of another machine you own as a Sona compute node
• Run the small Sona node agent there; every request from Sona is signed, and replays or tampered requests are refused
• The node only answers chat and read-only calls — nothing can pull, delete or change its models — and you can limit which models it offers
• Your nodes are yours alone: other accounts can't see, test, pick or be routed to them; Sona Auto can use them for you

2.1.0 — Cross-device sync

• Cross-device sync: chats you start, rename or delete on one device show up on your others within about 15 seconds — no reload
• An open chat refreshes by itself when a reply finishes on another device
• Unsent drafts follow you: start typing on your phone, finish on your desktop (never overwrites what you're typing)
• Settings changed on one device apply on the others

2.0.0 — Sona on every device

• Install Sona as an app from the website on iPhone, Android and desktop browsers — no app store (sidebar → Install app)
• Touch-friendly on phones: bigger tap targets, no zoom-in on text fields, layouts down to 320 px wide
• Opens instantly and even offline: the app is cached on your device; your chats stay on the Sona server and are never cached
• Continue where you left off: open Sona on another device and it offers the chat you were just in; returning to the app refreshes it
• Settings → Privacy → Signed-in devices: see every device and sign out one or all others; updates install with one tap

1.9.5 — Stability check before 2.0

• Everything new since 1.7 re-checked for privacy: arena runs, workflows and screen snapshots stay with your account
• Backups restore cleanly (tested on a copy of the live database); Sona stays fast with chats, an arena run and a workflow at once
• Image library updated to fix known security issues

1.9.0 — Project Echo — Sona Live

• Sona Live: the hands-free voice conversation can now see your screen — share a screen, window or tab only when you choose
• Each turn you send includes one snapshot of the shared screen (downscaled); nothing is recorded or sent in between
• The eye button asks about the screen right away; a clear banner shows while you're sharing, and stopping is one tap
• If your model can't see images, Sona Auto picks a local vision model for screen turns
• Snapshots are tagged, deleted automatically after 7 days (Settings → Privacy), or all at once with Delete all now

1.8.0 — Autonomous tasks & workflows

• Tasks in the sidebar: save workflows of up to 10 steps — chat, web search, the agent or image generation — each with its own model if you like
• Steps pass results along with {{previous}}, {{step1}}, {{input}} and {{date}}
• Run now, or on a schedule (every N minutes, daily or weekly) — on the server, even with Sona closed
• Every run becomes a chat you can read, continue or branch; new results show a badge on Tasks
• Unattended runs never act without you: approval-needing agent actions are skipped and reported, the browser tool is off, paid models only if you allowed them for that workflow

1.7.0 — Model Arena

• Model Arena in the sidebar: one prompt, two to four models answering side by side as they stream
• Blind mode hides which model wrote which answer (and shuffles them) until you vote
• Pick the best answer, call a tie or mark all as not good; your votes build a private Elo leaderboard
• Every answer shows its measured first-token time, speed and cost; models on the same GPU take turns so their speeds stay fair
• Continue in chat turns any answer into a normal conversation; paid models follow your paid-model rules and budget

1.6.5 — Reasoning and run insights

• A Think button next to the model picker sets how hard a reasoning model thinks: Auto, Off, Low, Medium or High (shown only for models that can)
• The model's reasoning streams in a collapsed line above the answer and stays out of the answer itself; Claude and ChatGPT show their providers' summaries only
• A live status line while you wait (“Waiting for qwen3 · 3.2s”, “Thinking”), and a Timeline tab for every reply
• Agent commands keep stdout and stderr apart, with the exit code
• Reasoning tokens and thinking time in the stats; detailed, redacted debugging for the owner account only

1.6.4 — Better image tools

• Image options under the message box: your GPU (free) or an optional paid image API, aspect ratio, size, quality, and up to 4 images
• Edit, variations, inpainting with a mask painter, and ×2/×4 upscaling on your GPU
• Reference images and optional prompt enhancement by a local model
• Paid image APIs (OpenRouter, OpenAI) show the price and ask every time, with a per-request limit; never a silent fallback
• ChatGPT and Claude plans can't make images in other apps, so they aren't offered

1.6.3 — Claude Code with your Claude plan

• Hand coding tasks to the official Claude Code (tool menu → Claude Code, or /claude), signed in with your own Claude plan
• Choose per task: your Claude plan allowance or API credits with a per-task cap
• Claude models in chat now use the native Anthropic API with your Console key
• Claude Code status, sign-in and billing in Settings → Agent; Sona never enables overages or auto-reload

1.6.2 — ChatGPT plan (preview)

• Connect your ChatGPT Plus or Pro plan with OpenAI's “Sign in with ChatGPT” (preview): sign in on your PC, then pick a ChatGPT model in the Cloud tab. Sona Auto never uses your plan on its own.
• Settings → Providers shows the connected account, links to Manage usage, and signs out by revoking access at OpenAI.

1.6.1 — Clearer cloud errors

• When your OpenRouter privacy settings exclude a model, Sona now says so instead of showing a cryptic error.
• Anthropic Claude keys are sent with Anthropic's own sign-in headers.

1.6.0 — Your two GPUs, and more providers

• Sona now uses both of your Ollama servers: the anxlab server and your desktop's RX 9060 XT over Tailscale. The model picker has a tab for each, plus one for cloud models.
• Sona Auto routes between them: quick questions go to a small loaded model, harder ones to the desktop's bigger models — and it steps aside while ComfyUI is drawing an image or a model would not fit in free GPU memory.
• Connect OpenAI, Anthropic Claude, Google Gemini, Groq or DeepSeek with your own API key, or OpenRouter with “Sign in with OpenRouter”. Paid models stay off until you allow them, and never go past your monthly budget.
• Settings → Providers can test every connection and benchmark models on your GPUs. Ollama cloud models (they run on ollama.com) are labelled and never count as private.

1.5.0 — Faster, steadier, better desktop

• The System Monitor updates about four times faster, and Sona Auto no longer pauses to refresh the cloud model list — it refreshes in the background.
• Image generation heals itself: if ComfyUI stops or crashes while idle, the GPU worker starts it again within about two minutes.
• Desktop app: "Start with Windows" in the tray menu (off unless you tick it; Sona then starts in the tray), and it now checks for updates every few hours while running.

1.4.0 — Smarter Sona Auto

• Secrets stay local: if a message looks like it contains a key, token, password or card number, Sona Auto only uses local models, whatever your routing policy.
• GPU-aware: while another reply is generating, Auto keeps the model that is already loaded instead of swapping it out — and says so. Models too big for the GPU are avoided.
• If the model Auto picked fails before answering, it retries once with the next model it was allowed to use, and tells you.
• Set a monthly budget for paid models in Settings → Privacy; Auto stops choosing them when it's used up.

1.3.1 — Steadier agents

• Fixed: after an agent task, the next reply could start by repeating Sona's internal record of the earlier run.
• When an agent describes the work instead of doing it, Sona now asks it once to actually use its tools — small local models finish tasks far more often.
• A run that hits its step limit no longer claims to have finished steps it never ran, and Continue asks for the remaining steps explicitly.

1.3.0 — AnxLabs connections

• Sona can see your AnxLabs systems: the System Monitor lists your AnxOS Control Center nodes and Agent Office with live status.
• In agent mode, ask about them ("is my Minecraft server running?") — Sona reads the answer from AnxOS and Agent Office themselves.
• Read-only by design: Sona never starts, stops or changes anything there, and only the administrator account can look.

1.2.0 — Memory, skills and continuity

• Sona now brings up the memories that matter for your question first, instead of only the newest ones.
• Skills: save how you like a recurring task done and run it in any chat with /name — as a chat, a web search or an agent task. Create them under Assistants → Skills; type / to see them.
• Agent runs remember what they did: a later message in the chat knows which steps ran, which failed and which files changed.
• A run that stopped at its step or time limit shows Continue, which picks up exactly where it left off. The step limit now also counts actions the model asks for all at once.

1.1.0 — Agents that browse

• Agent mode can use a live web browser: it opens pages, reads them, clicks and fills in fields — and asks you before every page it opens and every click or entry. Turn it on in Settings → Agent.
• The browser runs in its own isolated container and can only reach public websites; it never types passwords. Each step shows a screenshot of the page.
• More reliable agents: garbled tool calls are explained back to the model, an action that keeps failing makes it change course, and a dropped connection to the model is retried once.
• Several changes to one file can be made in a single step, all or nothing.

1.0.0 — Sona 1.0 (private)

• Sona 1.0 is the first stable private build: everything from the 0.9.5 release candidate, re-tested for regressions, security and recovery.
• Every part of the server now has an automated check that it refuses visitors without a sign-in and keeps one account out of another's chats, files, memories, projects and assistants.
• A restore drill brought a live backup back up in a separate container with every chat, message and file intact.
• Image generation recovers by itself if the graphics card stops responding: the GPU worker restarts ComfyUI and retries the picture once.

0.9.5 — Release candidate

• Sona 0.9.5 is the release candidate for 1.0: chat, memory, agents, Sona Auto, media, voice, updates and data export were tested end to end.
• An acceptance check script now runs the whole tour against a Sona server in about two minutes.
• Windows app: fixed a second, native title bar that could appear above Sona's own after updating from an early build.

0.9.4 — Faster and smoother

• Sona starts with about a third less code to load; settings, the Library, the System Monitor and code highlighting load when you open them.
• Replies stream more smoothly, especially in long chats: only the reply being written re-renders.
• The server no longer pauses other requests while it opens connections to models and tools.

0.9.3 — What's new

• Release notes like these appear once after each update — choose “Got it” and they stay out of your way.
• Open What's new in the sidebar any time for the full release history; a small dot means there is something you haven't read.

0.9.2 — Desktop updates

• The Windows and Linux apps check for updates after you sign in, and from the tray menu.
• Every update is signed by AnxLabs and must match the version it announces; nothing installs without your OK.
• Updates replace only the app itself — your sign-in and settings stay.

0.9.1 — Stability

• Stopping an agent run mid-approval is now recorded in its activity and the audit log.
• Security fix: the agent's undo history can no longer be altered from the sandbox.
• Reopening a chat whose model is gone keeps a working model selected.

0.9.0 — Sona Auto

• Sona Auto picks a model for each message — Local First, Balanced, Best Quality, Budget Saver or Privacy Lock.
• Each reply shows why its model was chosen; you can still pick any model yourself.
• Paid models are used only when you allow them in Settings → Privacy.

0.8.0 — Agent mode

• Choose Agent in the tool menu and Sona works step by step on files in the chat's own workspace.
• Commands run in an isolated sandbox without network access and ask for your approval first.
• Every step is shown live, can be stopped, and every run can be undone.

0.7.0 — Memory & context

• Sona remembers what you ask it to (“remember that …”) across chats — view, edit or delete it in Settings → Memory.
• Long chats are condensed automatically when they near the model's context window; your messages are never changed.

0.6.0 — Desktop apps

• Sona for Windows and Linux, with AnxLabs sign-in, a tray icon and a branded title bar on Windows.

0.4.0 — Voice

• Talk with Sona hands-free from the composer, with local speech recognition and Kokoro voices.

0.3.0 — Tools

• Web search with citations, image generation and editing on the GPU, one Library for everything, and the System Monitor.