Minimum hardware requirements

Lara locally runs an LLM (Ollama), voice synthesis with cloning (Coqui XTTS-v2) and, optionally, image/video/music generation (ComfyUI, MusicGen). Here's what you'll need, depending on how you use it.

What's New — v4.47.5

• v4.47.5 — Fix: the reference nginx configuration had drifted from the real server; resynced, Arabic locale blocking hardened (hard 404 across all 4 group sites).
• v4.47.0 — New: the app's version number is now shown under Settings → About.
• v4.46.9 — Security fix: replaced the Chromecast library (a "high" severity flaw with no non-breaking fix available) with a cleaner, better-maintained alternative.
• v4.46.8 — Security fix: Electron updated (2 high-severity security flaws fixed).
• v4.46.7 — Security fix: patched a critical RCE vulnerability (protobufjs) pulled in by the Chromecast feature's dependency — an npm override was silently failing to apply, fixed by deduping onto the already-patched version. npm audit: 1 critical → 0.
• v4.46.6 — Fix: corrected agent color themes (several no longer matched their avatar's actual colors) and regenerated Kanna's avatar; default app theme switched from "imperial" to "arctic".
• v4.46.5 — Fix: the logo's background made transparent — it used to sit on an opaque black tile, now blends directly into any surface.
• v4.46.4 — Fix: repaired an incomplete local Desktop installation (missing dependencies) — no functional change for end users.
• v4.46.3 — Fix: last remnants of gold branding on Android, iOS and the Linux phone client — icons, default theme and a few hardcoded colors repointed to imperial blue.
• v4.46.2 — Fix: the Desktop app's update check never actually worked (code never loaded) — replaced with a real check, clickable notification when a new version is available.
• v4.46.1 — Fix: last gold remnants on the desktop app — the Windows icon (taskbar, installer) still showed the old gold "L", regenerated from the new imperial-blue seal; two hardcoded colors (fallback avatar, ambient particles) fixed.
• v4.46.0 — New: TV casting to a DLNA television from the Linux phone client (Lara only for now).
• v4.45.0 — New: AirPlay TV casting from the iOS app (button in Settings).
• v4.44.0 — New: Chromecast TV casting from the Desktop — real mirroring of your screen (live transcoding).
• v4.43.0 — New: screen sharing with Lara via Remote Access, in addition to local screen sharing with the AI.
• v4.42.1 — Fix: several agents' descriptions, badges and statuses stayed displayed in French regardless of the language chosen in the desktop app; links between the group's sites that sometimes led back to the French version instead of the language being viewed.
• v4.42.0 — New: live screen sharing with the AI on the Desktop, explicitly enabled in Settings with confirmation on every start.
• v4.41.0 — New: "Cast screen" button in the Android app's Settings, the first step of a TV-casting feature.
• v4.40.1 — Fix: the Android startup crash persisted after the build-pipeline fix (v4.40.0) — a distinct root cause was found. The local vault (Libsodium encryption) needs a cryptographic capability missing from the mobile app's JS engine; fixed by adding a polyfill (Android and iOS), verified with a real emulator launch including a full account creation, no crash.
• v4.40.0 — New: shake your phone to report a bug straight to the developers, form available in all 7 languages of the mobile app. Fixed the server-side nightly build pipeline: Android and Windows builds had been failing silently for weeks (the native Android folder was never regenerated after switching to a direct Git clone, and the Electron version wasn't pinned) — only the Linux phone build kept getting published. The APK offered for download on this site had been stuck on a month-old version, known to crash on startup on recent Android devices.
• v4.34.0 — New FAQ page on the site, in 6 languages, answering common questions before you download. The site's mobile menu now works even if JavaScript is blocked by a privacy extension. On the app side: fixed the Arabic language code on mobile (dz→ar, inconsistent with desktop and able to break language sync from the PC), added translation infrastructure to the Linux phone client (previously unlocalized) in all 7 languages, and translated the ~154 remaining hardcoded French strings on Android/iOS's Activity, Chat, History, Missions, Personality, Settings and Update screens — technical debt documented for several weeks, now cleared.
• v4.33.0 — Your conversations come to life: AI-generated titles, a summary kept fresh with every exchange, and an at-a-glance view of the content created, files shared and emails sent in each one. Every creative session (Image, Video, Code, 3D, Studio Chloé) is now saved and always there when you come back to it. Copy and paste finally works everywhere, including on generated images. And two brand-new dedicated spaces have arrived: Studio Pedro for video, Studio Jinwoo for music and sound.
• v4.32.0 — Mobile mission execution (approve/reject/retry/cancel) on Android, iOS and Linux phone, plus 3 iOS fixes. Mobile Mission screens could create/edit/delete a mission but not drive its execution — those actions only existed on desktop since v4.12.0. Added on all three mobile apps with the same visibility rules as desktop (approve/reject only while awaiting approval, retry only after an error, cancel while still in flight). Also, an Android-vs-iOS parity audit found and fixed 3 real iOS bugs: a missing function that crashed every history action, a dead settings field that silently disabled the remote-access-request alert, and a version-check comparison that could misjudge once version segments reach two digits.
• v4.31.1 — Miscellaneous fixes: 5 missing full-body avatars now generated (Jinwoo, John, Oksana, Pedro, Yao — previously a broken image in the app); fixed broken navigation on all 6 Dutch pages of the site, which linked back to the French pages instead of their own translated versions; this page's download links now come from a single JSON manifest instead of being hardcoded across each of the 6 languages, to stop the recurring version drift.
• v4.31.0 — Remote Access: two-factor verification (2FA) and recognized devices. Once 2FA is enabled on the Remote Access account, a browser that has never cleared a code (or a passkey) gets a non-persistent session and must enter a 6-digit code before it can remember the access. New "Security" screen (from the remote chat): 2FA status, list of recognized devices (revocable individually), and full connection history — who connected, when, and from which device. Disabling 2FA automatically revokes every recognized device.
• v4.30.0 — The generated-content safety filter (image/video) was missing two ComfyUI node schemas (high-resolution SDXL and FLUX — the model used by default for automatic generation), which passed straight through unscanned: fixed. The "Free Mode" button (unrestricted adult content) has been removed: the filter is now always active, with no way to turn it off, and shows a clear error message pointing to the terms of use when it blocks a request. Deleting content from Lara's gallery now also deletes the matching original file from ComfyUI's own output folder.
• v4.29.1 — Added a minimum hardware specs sheet (documentation): required CPU/RAM/GPU/storage for text chat only, plus recommendations for local image/video/voice/music generation — new SYSTEM-REQUIREMENTS.md page and a table added to the download page, across all 6 site languages.
• v4.29.0 — Stronger family account controls: a member's access (allowed agents, quota, bandwidth) can now be edited after creation, not just fixed once and for all. Any outbound communication an agent triggers automatically (email, call) now waits for your approval before actually going out, unless it's to a contact you've personally confirmed. New dedicated email address for Lara so you can write to her directly. The primary account now has full access from their phone, on par with desktop. Stronger privacy: other family members' conversations stay private by default. The generated-content safety filter (image/video) now lives directly inside the app, for systematic protection independent of external services.
• v4.28.0 — Family account management (Settings → primary account): creating access for a family member is no longer limited to a fixed preset. Choosing "Custom" now lets the primary account precisely set which agents are allowed, the daily quota (in minutes), and the allocated bandwidth share, directly in the creation form.
• v4.27.0 — Website: the navigation bar (8 sections + language switcher) overflowed and broke onto a second line at common screen widths (1366/1440/1536px), with the mobile menu only kicking in below 768px — fixed (switches to the mobile menu at 1680px). Nav links now behave like real buttons (gold background on hover/active, instead of a plain underline). New scroll-reveal animation: content blocks (cards, team sheets, media) now fade in as you scroll down the page, instead of all appearing at once on load.
• v4.26.0 — Remote Access: security hardening and passkey login. The encrypted link between your browser and your Lara is now authenticated end-to-end (Ed25519-signed key + a verification code shown on both sides) — the relay can no longer sit in the middle. Remote chat can no longer trigger local tools on your PC. Stronger pairing codes. New passkey login (passwordless, phishing-resistant) on the remote access account, with the password demoted to a recovery method.
• v4.25.0 — Full front-end audit (desktop, website, mobile). The VoIP callback (phone fallback) had been broken since its introduction in v4.0: the SIP library wasn't even installed and the Electron sandbox blocked it from loading — fixed (library bundled locally, real authentication). The mobile app used to check the access code only locally on the device; it now authenticates that first login against the Desktop server (password + 2FA if enabled), then uses the local access code afterwards for offline unlocking — this also fixes a bug that prevented accounts other than Franco's from logging in. Stronger CSP across the site's 38 pages, favicon/robots.txt/sitemap added. Various fixes: garbled characters in the desktop assistant, obsolete files removed from the mobile repo, a small animation leak. ~240 dynamic desktop assistant messages (statuses, dialogs, errors) translated into the site's 5 other languages.
• v4.24.1 — The Linux phone client (Ubuntu Touch/UBports) finally builds successfully: fixed two Clickable configuration bugs (icon install conflict, outdated AppArmor policy version) that had blocked any build since the scaffold was added in v4.14.0. The nightly pipeline now also builds and publishes this client automatically. First .click package available for download above — still an experimental build, never tested on a real Ubuntu Touch device.
• v4.24.0 — New paid "Remote Access" option: chat with your personal Lara from any browser (Remote Access page), disabled by default — Lara stays 100% local until you explicitly enable it in Settings. A cloud relay connects your browser and your PC without ever seeing your messages in plaintext (end-to-end encryption, the same cryptography as the rest of the app); on the browser side, no persistent trace is left by default (session token kept in tab memory only, purged on close), with an explicit checkbox to remember the device on your own PC.
• v4.23.0 — Missions screen: real-time visibility into what the agent is doing while a step runs. Until now, between a step starting and finishing — which can take several minutes — nothing was shown beyond a generic "in progress" indicator. The mission now shows live which tool it's currently using, whether it's waiting on a result, or drafting its response, both on the mission card and in its detail view. Token accounting fix: missions, autonomous agent tasks and agent-to-agent delegations (e.g. Lara delegating to Yao) were generating real GPU load but were never counted in usage stats — fixed at the source, in the shared Ollama call function used across the whole codebase, so these stats now reflect actual total usage instead of just chat conversations.
• v4.22.0 — New « Directive + Continuous Mission » mode in Missions: combines a permanent directive (injected into every conversation) with a real, time-unlimited background task — the first plan requires approval like a classic mission, subsequent cycles are then automatically re-planned and executed without re-approval each time, until paused or deleted. Website translated into English and Italian (6 languages total: FR/CS/DE/NL/IT/EN). Accuracy fix: the Team, Agents and Creation Studio pages listed several outdated agent names or wrong job titles (leftovers from old renames never propagated to the site) — corrected across all 5 existing languages to reflect the 14 real agents.
• v4.21.1 — Fix: local music generation (MusicGen) could report an "unavailable" error when the server was simply still loading the model onto the GPU — the auto-start timeout increased from 3 to 5 minutes.
• v4.21.0 — End-to-end encryption of request bodies between the mobile apps and Desktop (on top of the signature already in place): a key unique to each device now also encrypts the content of API exchanges, not just their authenticity — useful notably on local networks. No existing app needs to be rebuilt, the upgrade happens automatically. Hardening of password/key encryption (Argon2id) with higher costs for the account and the database key. Cleanup: removed unused code inherited from a Capacitor experiment never used by the Lara mobile apps.
• v4.20.1 — Packaging fix: the Android APK published by the nightly pipeline still carried the version name "v3.7.0" (read from the wrong version file by mistake), instead of the app's actual version — which had been breaking the site's APK download link for a while. The Setup/Portable links also pointed to filenames different from those actually produced by the build. Both fixed.
• v4.20.0 — Windows only: real resume after an outage — an autonomous or Volt task still running when the PC closes/loses power is automatically relaunched on your next login (instead of just being marked as failed), and Volt tasks (previously lost without a trace on any restart) are now saved as well. New "Game Mode" in the tray menu: check it before launching a game to pause missions/Volt/autonomous tasks (between two steps, never mid-generation), uncheck it to resume.
• v4.19.0 — Windows only (Linux/Android packaging unchanged for this version): Lara now keeps running in the background (icon in the notification area) instead of closing completely, with a "Start with Windows" option; the daily watch now only triggers after login (it used to fail silently if launched before the database was unlocked); autonomous agent tasks now survive an app restart (shown as interrupted rather than lost); new "Volt Tasks" screen in Missions to approve/reject internal delegations without going through chat.
• v4.18.0 — New mode selector (Manual / Auto / Plan / Edit) on the conversation screen dedicated to Yao: Auto delegates the request to a multi-step autonomous task executed without approval, Plan has the internal planner produce a plan before any execution, Edit steers the agent toward concrete file modification. Related fix: submitting an autonomous task from the interface now correctly targets the chosen agent instead of always falling back to Lara.
• v4.17.0 — Creation/Download/Features/Team pages translated into Czech, German and Dutch (the site and app go from FR/EN/IT to 6 languages). Added a "Powered by Larai" mention in the footer of every site page. Internal fixes: autonomous agent tasks were in fact never invoking Ollama (missing call restored), and four pairs of legacy tool tags (agents renamed since) that could leak raw text into chat instead of being processed as a tool call.
• v4.16.1 — Additions to the mobile API audit: device signature now also covers the query string (an attacker on the local network can no longer alter the ?archived=/?since= parameters), timestamp freshness verification now fails closed if the received value is invalid, and the missing server-side conversation deletion route was added.
• v4.16.0 — Mobile API security hardening: device signatures now also cover the request body with anti-replay protection (previously: only the method and path were signed) — requires rebuilding the Android/iOS apps. Pairing code generation switched to a secure cryptographic generator. Strengthened access control on conversations and several admin routes.
• v4.15.0 — New enriched Cloud Services screen: real OAuth connection to GitHub (repos, issues), Azure (Blob Storage container listing), Google Calendar and Outlook Calendar (the existing calendar tool automatically switches to the real connected calendar, otherwise stays local as before). New capabilities Lara can use on demand in conversation: advanced email management (search, delete, mark read, drafts), and developer tools — list/start/stop/restart Docker containers and view their logs, open a file or folder in VS Code (no arbitrary Docker command is allowed, for security). Added an iCloud button in Cloud Services, which clearly states that Apple does not offer a public API for this service rather than failing silently.
• v4.14.0 — Linux phone client (Ubuntu Touch/UBports): added Missions screens (list, creation, editing, activation, deletion) and Personality (Modelfile editing + model rebuild + RAG memory stats), ported from the Android app. This client remains an untested scaffold for lack of an Ubuntu Touch build environment — see linux/README.md.
• v4.13.1 — The Android/iOS avatars had stayed stuck on the old style (illustration/sticker) instead of the photorealistic style used on Desktop since late June; resynchronized (Chloé, Durga, James, Karl, Lara, Laure, Mike, Sandro, Yao). The Yao avatar file mistakenly added on Android was a copy of Sandro's portrait, replaced with the correct one. The Durga avatar was missing on iOS (no avatar displayed for her messages), fixed. Leyla's avatar used an extreme auto-generated face crop instead of a bust portrait like the other agents, replaced with a photo from the validated batch.
• v4.13.0 — New profile analysis capability: Lara can scan a folder (local disk or authorized network share) and read files one by one to propose a more accurate personality profile, with progress tracking in the Missions screen and explicit approval before applying the result.
• v4.12.1 — The mission priority "Critical" fix (v4.11.3) had only been applied on the desktop side: the Android and iOS mobile apps were still sending a value not recognized by the server, silently falling back to "Normal". Fixed on both apps.
• v4.12.0 — Missions can now be real executed tasks: an agent builds a step-by-step plan, executes it with its tools, writes a final report in its own voice, and automatically resumes from the last completed step if the computer shuts down along the way. New tracking screen accessible by clicking a mission (plan, real-time progress, approval, rejection, relaunch).
• v4.11.3 — Missions screen audit: active directives were in fact never being passed to agents despite the "injected into every conversation" label (fixed, now actually injected, for all agents); mission creation/editing/deletion by any account (not just the main one) is now blocked for secondary accounts on both desktop and mobile; the "Critical" priority that matched no value recognized by the server and silently fell back to "Normal" was fixed.
• v4.11.2 — Another case of the "I am Lara" bug fixed: in an agent's dedicated conversation screen, clicking on another agent while a response was still being generated displayed that response (and its identity) under the name of the newly opened agent. The response is now correctly attributed to the agent that actually generated it.
• v4.11.1 — New case of the "I am Lara" bug fixed: in the main chat panel, switching agents mid-conversation (switch/join/leave) made it inherit previous turns from another agent with no author label, which could make it respond with the identity of the last agent who had spoken. Each turn now carries its agent's name.
• v4.11.0 — Volt missions and autonomous agent tasks now display real-time progress (%, current step) in the My Agents panel, and trigger a notification (as well as a mobile event) once the work is finished or has failed.
• v4.10.3 — Full security audit: "main account" authorization is now checked at tool execution time (and no longer merely suggested to the model), file explorer access confined to D:\, email/web page content treated as untrusted data, local learning storage that no longer ever writes in plaintext, and various hardening fixes to infrastructure scripts.
• v4.10.2 — 👍/👎 buttons under responses now always visible (instead of only appearing on hover) on all conversation screens — main chat, group chat, dedicated agent screens. An actual feedback submission failure is now clearly displayed instead of a fake success message.
• v4.10.1 — Important fix: every agent (Yao, Chloé, Leyla...) would sometimes reply "I am Lara" instead of its own identity — its persona was silently overwritten on every message. Fixed. Leyla's avatar replaced (wrong generated photo) and a bullet-list display bug in agent chat fixed.
• v4.10.0 — The Cloud Services screen now also connects to Gmail and Outlook: Lara can read your inbox and send emails on your behalf from your real account, with the same Connect/Disconnect buttons as for Drive/OneDrive/Dropbox.
• v4.9.0 — The Volt worker pool now runs on the real VoltAgent framework, still 100% local via Ollama: detailed plan, mandatory approval then tool-equipped execution (files, web search, visual creation delegation) by a dedicated agent.
• v4.8.0 — Self-learning: 👍/👎 buttons under each response let you flag an error (with an optional comment). The AI locally extracts a corrective lesson from it, reinforces it if the error recurs, forgets it if it doesn't recur again, and progressively builds a learned profile of the user. Relevant lessons are reinjected into context on every message (weighted semantic retrieval). Everything is local and encrypted; new "Autonomous Learning" panel in Personality to view, refresh or delete lessons.
• v4.7.0 — Context window extended to 256,000 tokens (from 65,536) for Lara and the 16 agents — much longer conversations and documents held in memory without premature compaction.
• v4.6.0 — Increased reliability of agent execution: file search <FS_SEARCH> now accepts a root (including the NAS \\nasfm32\AI) instead of being limited to D:\, NAS access errors become explicit and fixable, file tool failures are logged, and the 16 agents receive a "100% execution" directive (never a simulated result, retry on error, never a silent abandonment).
• v4.5.1 — Security fixes from an audit: path traversal in the mobile approval API, extension allowlist for opening system files, stored XSS in file history, confirmation before shutting down/restarting the PC.
• v4.5.0 — Lara and all agents now have direct read/write access to the NAS share \\nasfm32\ (SMB) without going through PIN unlock, which until now was reserved for the rest of the disk outside D:\. Deletion/moving on the NAS remains subject to the PIN as a precaution.
• v4.4.0 — New dedicated "Cloud Services" screen to link Lara to Google Drive, Microsoft OneDrive and Dropbox (connection, browsing, disconnection), accessible from Tools & Systems.
• v4.3.0 — Harmonization of all agent Modelfiles on the same qwen2.5:latest base (Yao, Durga, Chloé, Leyla, Hermes and Openclaw no longer went through an intermediate "lara" tier), with the same extended context (65536 tokens) and unlimited generation for all 16 agents. Cleanup of a dead Modelfile and an orphaned Ollama model left over from earlier renamings.
• v4.2.0 — Removal of artificial conversation limits: no more character cap on messages/forms (Lara conversation, agents, generation prompts, missions, projects), no more fixed timeout on Ollama/NIM responses (the "Ollama timeout exceeded" error should no longer appear), and delegated agent reports are no longer truncated before being relayed to Lara.
• v4.1.4 — Maximum message size multiplied by 5: 4000 → 20000 characters for the conversation with Lara, 3000 → 15000 for dedicated agent conversations.
• v4.1.3 — The "typing" indicator in agent chat now displays a phrase tied to each agent's function (Yao: "Auditing the code...", Laure: "Checking legal compliance...", etc.) instead of identical generic text for everyone.
• v4.1.2 — Strengthening of Yao's Modelfile (CTO): discipline of literal compliance with technical requirements (explicit constraints, requested format, zero invented APIs, performance criteria verified rather than merely asserted).
• v4.1.1 — Export button added to each agent's dedicated conversation screen (Yao, Leyla, etc.), previously only available in the main conversation with Lara.
• v4.1.0 — Explicit confirmation before any task delegation to Yao or Hermès. Finalization of the Wanda → Kanna and Julien → Karl renaming.
• v4.0.0 — New local TTS engine Coqui XTTS-v2 with voice cloning (GPU). Custom ElevenLabs voices for the 11 agents (Voice Design, a distinct voice per personality). Renaming Emma → Leyla (CTIO Translation). Twilio/SIP telephony configuration with built-in call test.
• v3.20.0 — Complete trilingual FR/IT/EN interface (~55 i18n keys). Trilingual agent badges with getAgentBadge helper. Kanna and Durga avatars loaded from Google Drive. Ennio removed from navigation.
• v3.19.0 — Floating agents menu, integrated file explorer, Cloud sync, context menus, full i18n. Privacy-safe activity logs with per-account filtering and security reports. Secure DoD 5220.22-M deletion (3 passes) on all files. Token stats auto-refresh every 60s with per-account detail. Positive/negative prompt fields visible in the image panel. Automatic test account creation on master login.
• v3.18.0 — Full-text search across conversations (magnifying glass in the history header, results with highlighting, click to open). History hidden outside chat view. Embedded OFL fonts: Noto Sans (body), Rajdhani (headings), Arya (logo). Unified typography across the whole interface.
• v3.17.0 — Oracle enabled for all agents. Tools organized by category (Internet, Generation, LLM Agents, Memory). On/off switches per LLM tool. Gold letter avatars. 52px agent photos in cards. 14 color themes per agent. Encrypted Franco profile injected into all agents.
• v3.16.0 — Collapsible navigation groups: Chloé Studio, Tools & Systems. Gallery folder management (create, rename, delete, move, copy, export).
• v3.15.0 — Agent first names capitalized. Restructured navigation. VRAM optimization: immediate release on close, secondary agents migrated to qwen2.5.
• v3.14.0 — Fullscreen body avatar for Lara: full-length shot, reinforced dark blonde. Main console ASCII encoding fixes.
• v3.13.0 — Brigitte & Julien enabled in the UI. Sandro: Safe by Design, GDPR, NIS2, mandatory AES-256/TLS encryption, agent training mission. Agent routing fixes (@mention only) and history display fixes.
• v3.12.0 — Brigitte (Head of Ethics & Philosophy): PhD in philosophy, specialist in applied ethics, social criticism, African philosophy (Ubuntu, Maât), AI ethics. Tag <BRIGITTE>.
• v3.11.0 — Julien (Head of HR): recruitment, dismissal, candidate selection. XLSX/Excel attachments, Vision screen capture, clipboard, Playwright browser. Ghibli/anime/Pixar/3D animation styles. Whisper STT + multimodal Vision.
• v3.10.0 — Gallery: full-size hover preview, multi-select + batch deletion, fullscreen. Universal sidebar. Generation history. Parental control. I: drive encryption. Universal Chloé delegation. Quantity 1-10 per generation. ComfyUI fixes: video duration respected, auto-relaunch after crash, generation stop, progress bar, AVI + 1080p/2K/4K
• v3.9.0 — Conversation export (MD/HTML/PDF/Word/ODT/JSON), gallery lightbox + rename + edit, NSFW Free Mode, model selector, Creation Studio page
• v3.8.0 — Creation Studio: cinema video (quality presets, 5-60s duration, fps, formats), 4K HiResFix photo, I/O token counter, statistics screen, avatar gallery import
• v3.7.0 — High-definition manga avatars for the 11 agents + team page on the website
• v3.6.0 — 7 new color themes (Hibiscus, Africa, Europe, Asia, Oceania, Americas) + UTF-8 encoding fix and hallucination fixes
• v3.5.0 — Laure promoted to CLO (Chief Legal Officer) with full lawyer missions and Belgian code of ethics
• v3.4.0 — Lisa (Head of AI Relations) + Scott (Lead Writer) + Sandro CTO + bidirectional communication between agents
• v3.3.0 — Volt Pool: autonomous workers with mandatory approval + Project Management screen
• v3.2.0 — Emma (CTIO Translation & Interpretation) + C-level / H-level org chart
• v3.1.0 — Ennio (music) + Steven (video) + complete hierarchical restructuring of agents
• v3.0.0 — MusicGen music generation (Python 3.11) + AnimateDiff video generation — new embedded AI models
• v2.0.0 — Ed25519 OTA updates, built-in WireGuard VPN, premium UI, website, iOS/Android mobile apps