InnerZero logoInnerZero
Download

Changelog

What's new in InnerZero.

v0.2.1August 2026Latest
New
  • Attach your app log to feedback: the Give feedback box in the sidebar can now include the last 15 minutes, hour, or day of InnerZero's own activity log. Only event names, labels, and sizes are sent; prompts, replies, file paths, and your username never leave the machine, and you see the exact preview before it goes. Settings > General also gained an Open Logs Folder row.
  • Ask Zero about InnerZero itself: questions like "how do I change the theme" or "what can you do" are now answered from a built-in guide to the app, and asking how to report a bug points you at the feedback box.
  • Use an Ollama server on your network from a low-spec machine: if your PC cannot run a local model, point Settings > AI & Models at another computer on your network running Ollama and chat works from there. The chat banner, blocked messages, and the setup wizard now offer three ways forward on that tier: a cloud plan or API key, choosing a tier yourself, or connecting a network Ollama server.
Improved
  • Bundled Ollama updated from 0.22 to 0.32. Vulkan GPU acceleration is now on by default, so AMD and Intel GPUs work out of the box and the old experimental Vulkan toggle in Settings is gone. NVIDIA GPUs need driver 550 or newer for GPU acceleration, or 570 or newer for older Maxwell and Pascal cards, which may also run a little slower than before; older drivers fall back to the CPU. On Windows, AMD cards on the ROCm path need ROCm 7 drivers, and RDNA2 cards are now served by Vulkan instead.
  • Office documents in chat: .pptx, .docx, .xlsx and other Office and document types are now recognised as files when you ask Zero to open, read, or find them.
  • Settings > Hardware now names every GPU it detects rather than just the first, and machines with 32 GB installed that report 30 to 31 GB usable now pass the RAM checks for their proper tier.
  • The first-run GPU check no longer fails on a cold first model load on slower machines.
  • Connecting a network Ollama server no longer suggests a restart that changes nothing: Save tells you to pick an assistant model, and the model list refreshes from the new server straight away.
  • Voice follows your chosen cloud source (managed plan or your own key) instead of quietly using the other one, and models InnerZero has not vetted are flagged in the voice and setup pickers as well as the assistant picker.
Fixed
  • A machine on the No Local Model tier no longer answers "hi" and then blocks every real question, and voice now follows the same tier gate as chat instead of answering a turn that chat would refuse.
  • Updating on Windows now removes the previous version's AI runtime libraries instead of leaving them beside the new ones.
  • Deleting a calendar event that was pushed to Google no longer loses its link to the Google copy before you confirm.
  • Unrestricted Mode: the manual model install instructions stay visible when you leave the Experimental tab and come back.
Known limitations
  • Machines with only an integrated GPU that used the old Vulkan toggle may run on the CPU after this update, because Ollama 0.32 hides integrated GPUs by default. Set the environment variable OLLAMA_IGPU_ENABLE=1 before starting InnerZero to use it again.
v0.2.0August 2026
New
  • High Contrast theme: a seventh free theme with a true-black background and maximum-contrast text on every surface, built for low-vision use. Alongside it, every theme was brought up to WCAG AA text contrast and all text below 11px was raised to a readable minimum.
  • Drag and drop files into chat: drop a document or image straight onto the chat box to attach it, exactly like using the paperclip.
  • Copy button on chat messages: copy any reply as clean plain text, without the hidden styling that could paste as invisible near-white text into other apps.
  • Setup can adopt models you already have: if a suitable model is already installed, the first-run wizard offers to use it for the assistant or voice role instead of forcing a fresh download. Choosing a tier and committing to it are now separate steps, so one click can no longer finish setup by accident.
  • Honest startup: the window now opens immediately and a progress line names the exact step the app is on, instead of blocking invisibly for up to half a minute on slower machines. The first-run wizard gained its own loading screen too.
  • Safer knowledge pack downloads: packs now check disk space first, download to a staging file, verify a checksum, show real megabyte progress, and can no longer be started twice at once and corrupt themselves.
  • Unvetted model warning: picking an assistant model InnerZero has never vetted for tool use now shows a clear note, instead of silently claiming everything is fine.
  • Better hardware detection: Windows now names AMD, Intel, and integrated GPUs instead of reporting none, and Linux machines with two AMD GPUs are detected and tiered correctly.
Improved
  • Marking a calendar event private now reliably removes the copy that was previously synced to Google, even if the connector is disconnected at the time; the cleanup completes when it reconnects.
  • Approvals got honest end to end: an approved action that then fails shows as a failure rather than a green tick, decided approval cards stay decided when you navigate away and back, and a reply from the model can no longer draw a working approval button of its own.
  • Unrestricted Mode reliability: the toggle no longer unticks itself for users whose consent was recorded under an older format, the off switch stays reachable even when the model is missing or the backend changed, the uncensored-model badges now light up in every case they should, and the mode is clearly disclosed as inactive on LM Studio and llama.cpp.
  • Costs page accuracy: all eight time period options now work, and the bring-your-own-key cost estimates were refreshed to provider prices verified in August 2026. One model had been under-stated by around twelve times. DeepSeek estimates use its new schedule's peak rate, so off-peak usage may be over-stated but is never under-stated. Old logged usage keeps the rates it was billed at.
  • Fewer wrong tool detours: the assistant no longer fires the calculator, date, or location tools on innocent words like "update" or "chicago", answers ordinary questions directly instead of preferring tools, and no longer returns raw encyclopedia dumps for simple questions.
  • Chat clears cleanly: clearing a conversation now takes effect everywhere at once, a reply that was still being written stops instead of painting into the cleared chat, and cleared history no longer reappears on your next message. Chat also now discloses its seven-day history retention and how many old rows were trimmed.
  • Status bar honesty: cloud availability is now a three-state signal instead of showing green until a send fails, cloud-only setups no longer show a permanent red local-AI warning, and the label names the engine that is actually down, including llama.cpp.
  • Startup recovery: if the local AI service fails to start, the cause is now recorded and the Retry button genuinely restarts it; the chat warmup gate also recovers by itself on slow machines instead of leaving Send blocked.
  • First-run polish: the disk space check now counts everything it lists, download bars show reconnection attempts instead of freezing, failed steps are reported as failures instead of a clean finish, re-running setup asks before changing models you already chose, and setup ends with a summary and a Continue button instead of a flash.
  • Voice: the start of an utterance is no longer clipped when speaker recognition is active, changing the voice model takes effect without a restart, and voice no longer reads out tool payloads or approval reference ids as if they were the answer.
  • Windows: files now save to your real Documents folder, including when OneDrive has redirected it, and in-app updates now guard against relaunching mid-install and show a banner if an update failed. macOS and Linux: the app now sees models in your existing Ollama store instead of forcing its own.
  • Screen readers now narrate the interface in your selected language, spell-check follows your locale instead of flagging everything as misspelt English, and countdown and wizard messages are translated across all 26 languages.
  • Error messages from the AI engine are no longer stored into memory, so a bad afternoon can no longer contaminate later answers. Tool failure messages no longer embed file paths or usernames.
Fixed
  • Navigating away from a page while it was still loading no longer paints the old page over the one you moved to. Closed on Chat, Agents, My Privacy, Costs, the dashboard, every Settings tab, and the knowledge pack downloader.
  • Leaving Chat and returning no longer discards a file you had just attached, and My Privacy can no longer briefly claim Private while the backend is still in Cloud mode.
  • The status bar clock no longer wipes what you are typing into dropdown fields once a second, so custom timers and labels can actually be set.
  • A reply sent from Telegram no longer types itself into the desktop chat box, and agent status messages no longer duplicate every time you reopen Chat.
  • Specialist agents: a failed specialist call is no longer presented as a successful answer, a crash restores your main assistant model so the next message works, and a specialist's prepared task brief is no longer silently thrown away on GPU and cloud setups.
  • Overnight memory processing no longer marks items as done when the extraction actually failed, and a malformed cloud response no longer permanently ends a document's retries. Degraded overnight runs are now visible in the status bar with the cause.
  • On LM Studio and llama.cpp, having no chat model selected now says exactly that instead of replying "could you rephrase".
  • Visiting the Schedule page twice no longer breaks calendar connector updates.
  • The in-app Help page was removed; Settings and the / command browser already cover everything it showed. The /help command still works.
v0.1.9July 2026
New
  • Refreshed the bring your own cloud model catalogue. If you connect your own provider API key, InnerZero now offers the current flagship and tier models for each provider: OpenAI GPT-5.6 and GPT-5.5, Anthropic Claude Opus 4.8, Sonnet 5 and Fable 5, Google Gemini 3.5 Flash, 3.1 Flash-Lite and 3.1 Pro, Qwen 3.7 Max and Plus, xAI Grok 4.5 and 4.3, Kimi K3, K2.6 and K2.7 Code, and DeepSeek V4 Pro. Retired the older xAI Grok model ids that no longer exist upstream.
Fixed
  • Bring your own OpenAI keys now work with GPT-5 and o-series models. Warmup and chat were failing with a request error on these models; InnerZero now sends the correct token limit with a reasoning-token allowance, so they respond normally. GPT-4o and GPT-4.1 keys were unaffected.
  • Build and packaging fixes so the Windows, macOS, and Linux installers all build and ship cleanly.
v0.1.8June 2026
New
  • llama.cpp engine support: run InnerZero against your own llama-server. Pick llama.cpp in Settings, point it at your server, and chat, voice, and memory all route through it, with a model picker that lists your GGUF models and shows their names correctly.
  • AI & Models settings tab: your AI engine, hardware, model assignments, performance, and cloud API keys now live together in one tab. Cloud mode and the Privacy Blacklist live on the My Privacy page, and plans on Plan & Usage, so the old Mode tab is gone.
  • Open at login: a new toggle in Settings starts InnerZero automatically when you sign in to your computer, on Windows, macOS, and Linux. Off by default, per-user only, and it never asks for admin rights.
  • Action Hub saved actions and scheduling: save a research or apply action once, re-run it in one click, or put it on a schedule. Scheduled runs always stop at a draft for your review, never auto-submit. The scrape history table also got a cleaner, expandable layout.
Improved
  • Settings layout: the General tab is shorter and better ordered, Auto Sleep moved next to Memory where it belongs, and model dropdowns are wider so long model names are readable.
  • Voice settings labels are now in plain language, in all 26 interface languages.
  • Model downloads now show the real error message when a download fails, instead of a generic unknown error.
Fixed
  • Sending a message while a file attachment was still being read no longer drops the message. It now sends automatically the moment the attachment is ready.
  • GGUF model names no longer lose their quantisation tags in menus, so you can tell model variants apart.
  • macOS open-at-login can no longer record a broken login entry when the app runs from a mounted disk image or before it is moved to Applications. The toggle now declines safely and reports its real state.
  • Every in-app hint that pointed at the old Mode tab now points at the right place.
v0.1.7June 2026
New
  • Multi-language interface: use InnerZero in 26 languages. Pick yours in Settings and the whole app, its menus, and the assistant's replies follow, with right-to-left support for Arabic, Hebrew, Persian, and Urdu. A one-time upgrade keeps your saved memories searchable across every language, and it all works offline.
  • Prompt Library on the Chat page: save your best prompts in folders, mark favourites, and drop any of them into chat with a single click.
  • Voice page panels: a standalone speech-to-text panel that turns what you say into editable text, and a text-to-speech panel that reads any typed or pasted text aloud, both running locally with keep-warm and stop controls.
  • Attach files and images to chat: add documents (.pdf, .docx, .xlsx, .csv, .txt, .md) or images with the paperclip, or paste a screenshot. Zero reads them, runs local text recognition on images, and answers over their content.
  • Artifacts and export: document-style answers open in a side panel you can read full screen and edit with the AI, then export to PDF, Word, Markdown, HTML, and more. PDFs keep real, selectable text.
  • Dictate in chat: press the microphone in the chat box and speak; your words appear as editable text before you send.
  • Action Hub: an opt-in research and apply assistant on the Tasks page. With your own Apify key, Zero can gather sources from the web; with your saved job profile, it can help fill in and submit job applications. Every action is gated behind your approval, and the apply browser runs in a fresh, isolated session.
  • Proxy support: behind a corporate or university proxy? Enter your HTTP or HTTPS proxy in Settings and InnerZero routes model downloads, cloud calls, and connector traffic through it. Local AI traffic is never proxied.
  • One-button first-run setup: a single, friendly setup screen with one progress bar covering the AI model, components, and memory upgrade, with clear states for offline or slow connections.
  • InnerZero Pro: an optional membership that unlocks premium themes (including Golden Pro with an animated starfield), premium personalities, and extra proactive briefing types. It works fully offline and is separate from cloud plans.
Improved
  • Privacy egress guard: every outbound connection now passes through a single fail-closed guard, so Offline and Private modes hold even if a tool tries to reach out.
  • Smaller, faster Windows install: dev-only tooling is excluded from the shipped runtime and the heavier components now download on first run.
  • Faster responses and startup: cloud engines and local model connections are reused across calls, memory search is cached within a session, and hardware details are prewarmed at launch so Settings opens instantly.
  • Action Hub now runs on macOS and Linux, not just Windows.
Fixed
  • macOS now opens cleanly. The v0.1.7 DMG is signed with Developer ID, notarised by Apple, and stapled, which fixes the "InnerZero is damaged and can't be opened" message some macOS users saw on v0.1.6. No right-click workaround is needed.
  • Project-scoped memory isolation: memories saved under one project no longer surface in another project's context.
  • Voice reliability: the text-to-speech button no longer hangs on a slow first load, and it now shows clear loading and stop states.
  • Calendar sync correctness: private events are reliably kept off Google, events you make private after syncing are removed from Google, and multi-day all-day events now cover the correct days.
  • Windows launch no longer briefly flashes a black console window.
  • Speaker recognition data now lives in your protected user data folder rather than the app install directory.
v0.1.6May 2026
New
  • Proactive Assistant: schedule briefings, daily summaries, and reminders that arrive when you want them. Natural-language scheduling like "remind me at 5pm tomorrow", quiet hours, multi-channel notifications, and an optional Telegram bridge if you want briefings on your phone.
  • Smart briefings: a higher-quality briefing mode that pulls from your memory, mail, and knowledge packs as needed instead of relying on a single prompt-and-response.
  • Automation Specialist: route automation tasks to a small local model, your main assistant, or a dedicated cloud model based on what you want to optimise for (speed, quality, or privacy).
  • Slash commands in chat: type /help to see what is available, /clear to reset the visible conversation, and more for quick actions without typing prompts.
  • First-run hardware detection: InnerZero now tests your GPU on first run and picks the right model tier automatically. You can also override it manually in the setup wizard.
  • Bundled Ollama upgraded to v0.22.1 with full GPU acceleration: modern NVIDIA via CUDA 13, Apple Silicon via Metal, AMD and cross-vendor via Vulkan, and Intel via OneAPI.
  • AMD GPU users can now opt in to GPU acceleration in Settings. AMD support is conservative-by-default for this release while we collect real-world feedback.
  • Single-instance lock: launching InnerZero a second time now focuses the existing window instead of opening a duplicate.
Improved
  • Chat reliability for LM Studio backends: better handling of replies that drift from the expected response shape, and replies that contain emoji or non-ASCII characters now render correctly on Windows.
  • Linux installer streamlined for modern NVIDIA (Turing onward) via CUDA 13, plus Vulkan for AMD and cross-vendor.
  • Mac code-signing hygiene: cleaner Gatekeeper experience on first launch.
Fixed
  • Bundled Ollama on Mac and Linux now resolves correctly. Users on previous versions whose bundled Ollama silently failed to start should see chat working out of the box on v0.1.6.
  • LM Studio chat replies containing emoji or smart quotes no longer crash on Windows.
  • LM Studio: when a model returns a response shape that does not match what InnerZero expects, the actual answer text is recovered where possible instead of showing a raw fallback.
  • Several internal coverage gaps in the installer bundle that affected slash commands and other modules in installed builds.
Known limitations
  • The macOS DMG ships unstapled in v0.1.6. The app is signed with Developer ID, runs under hardened runtime, and applies the necessary entitlements. On first launch with internet, Gatekeeper does a one-time online check with Apple and the app opens normally. If you are offline on first launch, right-click the app, select Open, then click Open again. Stapling returns in a future release.
  • macOS minimum is now Sonoma 14. Bundled Ollama 0.22.1 no longer supports Monterey or Ventura.
  • Pre-Turing NVIDIA Linux users on the bundled mode (GTX 9 series Maxwell, GTX 10 series Pascal, V100 Volta) should use Vulkan or install Ollama separately for full CUDA support.
  • Installer sizes have grown vs v0.1.5 because of the Ollama 0.22.1 bundling. See the download page for current sizes per platform.
v0.1.5April 2026
New
  • Calendar page with Month, Week, Day, and Agenda views. Click an empty slot to create an event. Drag events to reschedule.
  • Two-way Google Calendar sync. Connect your Google account in Settings; events from Google appear locally and events you create at home publish back. Private events stay on your machine.
  • Gmail integration (read-only). Sender, subject, and a short snippet of recent inbox emails so Zero can answer questions about your mail. Message bodies are never fetched or stored.
  • Tasks page with a live queue. Kanban lanes, progress bars, ETAs, and resource coordination so you can see what Zero is doing in the background.
  • Dashboard 7-day calendar widget with a next-up section.
  • AI agency over your calendar. Zero can create, find, update, and delete events directly from chat with approval gates on writes.
  • Source provenance on memories. Every memory now records where it came from (chat, voice, document, Gmail, calendar) and surfaces that source to the AI.
  • Mac code signing with Developer ID and hardened runtime. The macOS installer is now signed by Summers Solutions Ltd.
  • Windows installer hardened with Azure Trusted Signing and a deferred-swap auto-updater that resolves the WebView2 file-lock during in-app upgrades.
Improved
  • Anthropic prompt caching is now active on Director calls, reducing cost on repeated cache-hit prompts within the 5-minute window.
  • Calendar-aware memory. Time-sensitive questions surface upcoming 7-day events directly into Zero's context.
  • Knowledge pack search quality. Two-phase title boosting, prose extraction, and question-prefix stripping produce cleaner answers.
  • Faster voice shortcuts. Time, weather, calculator, dictionary, system info, and timer queries now respond in under two seconds.
  • Privacy hardening for cloud features. Every cloud dispatch path (initial messages, retries, and multi-round agent loops) now routes through one privacy-blacklist chokepoint.
Fixed
  • Windows in-app updater no longer fails on the WebView2 DLL file-lock. The new deferred-swap pattern applies the upgrade at next cold start.
  • Archive page now shows archived memories correctly.
  • Settings hover styling and tab consistency across all 9 tabs.
  • Mac and Linux launches are now reliable: pywebview backend bindings ship with the correct platform markers.
  • Memory system: sleep pipeline routes correctly on the LM Studio backend, preference-type memories reach the Director prompt, and project-scoped retrieval backfills bidirectionally.
v0.1.4April 2026
New
  • Claude Opus 4.7 support. Use it with your own Anthropic API key in chat or with the coding specialist.
  • Frontier model tier for datacenter-class hardware (256 GB+ RAM, 120 GB+ VRAM).
  • Enthusiast coding model tier for high-end workstations.
  • Four new coding models: Qwen3 Coder Next, DeepSeek Coder V2, Codestral, and CodeGemma.
  • Tier switch preview shows disk space required before you commit to the change.
  • Uninstall downloaded models individually, with protection on any model currently assigned to an active role.
  • Theme redeem codes now work reliably end to end.
Improved
  • Coding agent reliability on long-running tasks.
  • Model downloads auto-resume if your connection drops mid-way.
  • LM Studio voice model picker now has feature parity with the Ollama picker.
  • Cloud account connection has better error handling and retry behaviour.
  • Memory system correctness across projects and specialists.
Fixed
  • Specialist now connects correctly to remote Ollama servers on your network.
  • Coding model dropdown shows all installed compatible models.
  • Coding agent parser no longer strips code fences when writing markdown files.
  • Coding agent can read files it wrote earlier in the same run.
  • Removed duplicate strategy content in cloud prompts.
  • Preference memories reach the assistant again.
  • Working state no longer leaks between sessions.
  • Fact verification works on the LM Studio backend.
v0.1.3April 2026
New
  • AI Specialists: delegate coding tasks to a specialist AI agent. Full file review and approve/reject before any changes are applied.
  • LM Studio support: use LM Studio as an alternative local AI backend alongside Ollama. Switch instantly in Settings.
  • Offline mode: completely block all outbound network requests with a single toggle. Nothing leaves your machine.
  • Connection log: see every outbound request Zero makes, with destination, timing, and status.
  • Privacy blacklist: define sensitive terms that are automatically scrubbed from all cloud messages before they leave your machine.
  • My Privacy page: centralised privacy dashboard with mode selector, blacklist management, connection log, and data controls.
  • Telegram remote access: control Zero from your phone via a Telegram bot. Encrypted token storage, chat ID whitelisting, and desktop chat mirroring.
  • xAI Grok and Kimi (Moonshot) cloud providers. 7 cloud AI providers now supported.
  • Neon Tokyo exclusive theme: cyberpunk purple and cyan with animated synthwave perspective grid. Unlock with a founder code from Discord.
  • Theme unlock system: redeem exclusive codes for special themes.
  • Costs page with currency selector (7 currencies), period filters, and per-request cost breakdown.
  • Windows installer now signed by Summers Solutions Ltd via Azure Trusted Signing.
Improved
  • Cloud voice now offers Standard mode (split reasoning and TTS), roughly 15x cheaper than Premium mode.
  • Cloud token usage reduced by approximately 80% for chat messages through context optimisation.
  • Choose different cloud models for Director and Specialist roles independently.
  • Specialist agent memories are now processed separately during sleep, with dedicated fact extraction and cleanup.
  • Cloud billing is now idempotent. Retried requests after timeouts will not be double-charged.
  • Account tokens refresh automatically on expiry. No more manual re-login after sessions expire.
  • Settings page loads significantly faster with lazy tab loading.
v0.1.2April 2026
New
  • macOS support: .dmg installer with .app bundle (Intel and Apple Silicon)
  • Linux support: AppImage for x86_64 with bundled Python runtime
  • Auto-updater: checks for new versions on startup, one-click update with SHA256 verification
  • GPU detection for NVIDIA, AMD (ROCm + HSA override), Intel Arc (oneAPI), and Apple Silicon (Metal)
  • Vulkan toggle for GPU acceleration on non-NVIDIA hardware (experimental, manual only)
  • Ollama mode persistence: bundled or system Ollama config saved from first setup, prevents model-not-found errors
  • Discord community link in the sidebar
  • System dependency notices for macOS and Linux on first launch
  • GPU acceleration section in Settings with detected backend display
  • Model location info in Settings for debugging
  • CI/CD pipeline: GitHub Actions builds all three platforms in parallel on tag push
Improved
  • Linux AppImage reduced from 2.7 GB to 356 MB (torch CPU-only install in CI)
  • Setup wizard shows retry and skip buttons with troubleshooting guidance when model downloads fail
  • Settings shows "No compatible GPU detected, CPU mode will be used" with manual tier override note
  • Sleep subprocess uses python.exe instead of pythonw.exe for reliable .pyc execution
  • Auto-sleep defaults to off on new installs
  • Consent modal checkbox text alignment improved
Fixed
  • Model not found (404) error when using system Ollama alongside InnerZero
  • Sleep pipeline crash in installed app (load_dotenv .pyc incompatibility)
  • Sleep subprocess failing silently (pythonw.exe, missing PYTHONPATH)
  • Version display showing "?" in Settings before background check completes
  • Discord sidebar showing raw template literal instead of icon
  • fetch_url Unicode crash on Windows (arrow character in print statement)
  • Auto-sleep toggle snapping back to off when dropdown closes on click
v0.1.1April 2026
New
  • Remote Ollama support: connect to Ollama running on another machine on your network
  • Unrestricted Mode with full age verification and consent flow
  • Automated memory backup system (weekly, up to 10 backups)
  • Version update gate: older installs prompted to update on launch
  • Business licence validation for commercial users
  • Star ratings on AI responses (1-5 stars, replaces thumbs up/down)
  • Custom alarm sounds: pick your own audio file for alarms
  • Memory import: paste text or upload files to teach Zero new facts
Improved
  • Clock system redesigned: preset timer pills, phone-style AM/PM alarm picker
  • Settings reorganised from 11 tabs to 9 (cleaner, less overwhelming)
  • Memory page replaced by Settings Memory tab (Core Facts, Recent Memories, Archive)
  • Sleep progress estimates are more accurate and never increase during a run
  • Auto-sleep with configurable idle timer (15-60 minutes)
  • Wake buttons in chat and status bar during sleep
Fixed
  • Dictionary Unicode crash on Windows (IPA phonetic symbols)
  • Weather API updated (Open-Meteo deprecated old endpoint)
  • Voice name confusion (reordered confidence checks)
  • Voice math shortcuts ("seven times seven" now uses calculator, not the AI model)
v0.1.0April 2026
New
  • First public release of InnerZero
  • AI chat with streaming responses and cancel support
  • Full voice mode: speech recognition, natural TTS, voice shortcuts
  • 30+ built-in tools (web search, calculator, timers, notes, file operations, and more)
  • Persistent memory system with local storage
  • Sleep/reflection pipeline for overnight memory processing
  • Knowledge packs (offline Wikipedia)
  • 5 themes (Dark Zero, Light Zero, Classic Carbon, Soft Pink, Dark Teal)
  • Cloud mode with BYO API keys (DeepSeek, OpenAI, Anthropic, Google AI, Qwen)
  • Cloud voice (OpenAI Audio with 13 ChatGPT voices)
  • AI personality system (Professional, Friendly, Concise, or custom)
  • Screen automation (read screen, click, type, scroll other apps)
  • Document upload and Q&A (.txt, .md, .pdf, .docx, .xlsx, .csv)
  • Project system for organising work and scoping memory
  • Hardware auto-detection and model selection
  • Setup wizard with guided first-run experience
  • Chat session persistence across restarts