What's new
Every release of Agentic, newest first. The app checks for updates automatically and offers to download new versions from here.
Version 1.5
Latest · August 2026
Iterative builds on top of Version 1.4’s Agentic Browser. An experimental On-Device mode was available in builds 30–34 and has been removed in build 35 to keep the app simple. A refined on-device experience is planned for a future release.
- Builds 30–34 introduced an experimental On-Device option with downloadable models and local progress.
- Build 35 removes On-Device and returns the app to Local (Ollama) + cloud providers for a cleaner experience.
Build 35 — On-Device removed
- On-Device has been removed to keep the app simple and reliable.
- The app now uses Local (Ollama) and cloud providers. A refined on-device experience will return in a future release.
Builds 30–34 — Experimental On-Device (later removed)
- Introduced an experimental On-Device option with downloadable models and model management in Settings.
- Improved install reliability, memory-aware model selection, and clearer download progress.
Version 1.4
August 2026
The headline feature: a built-in browser the agent controls. Ask it to look something up, fill a form, or use a web app — you watch it click and type live inside Agentic's own browser window, with far more precision than screen control.
- Agentic Browser. A real browser window inside the app. The agent reads pages, clicks elements, and fills fields precisely — no screenshots, no guessing.
- Six browser tools. browser_open, browser_read, browser_click, browser_type, browser_scroll, and browser_back — the agent browses like a person would.
- Your logins work. Sign into sites inside the Agentic browser and the agent can use them; a clear-browser-data button is included.
- Questions with tappable answers. When Agentic needs a choice, it asks with suggested answers you can tap — or type your own.
- Responses without length caps. The model decides how long its answers are.
- Optional unlimited steps. A new setting lets the agent keep going until the task is done.
- New sounds. Fresh response, error, and cancelled chimes.
Build 29 — Browser history and native menu
- Browser history — global history with grouped view and search (Today/Yesterday).
- History menu — Recent items and full history with separate clears for history and browsing data.
- Native menu bar — File, Edit, View, History, and Window menus when the browser is active, plus a refined agent panel.
Build 28 — required updates (per-version, full block, offline)
- Required updates can be configured for specific versions — when marked as required, the app blocks until updated.
- Required banner is red, non-dismissible, with fixed message “This update is required to continue.” plus blocking sheet.
- Chat/composer, file attach, and drag-drop are disabled until updated; works offline too (cached required stays blocking).
- Alert has only
Update Now/Download (no Later) and re-announces if dismissed.
Build 27 — browser reliability & VoiceOver overhaul
Reliability
- Critical fix: tabs now have their browser engine attached — back/forward, URL, title, and downloads work instead of staying stuck on “New Tab”.
- Tabs share one browsing session, so logins and cookies persist.
- Per-tab element refs — switching tabs no longer shows stale click targets.
- Better loading detection with timeout so frozen pages don't hang the agent.
- Downloads auto-open the save panel; JavaScript dialogs no longer freeze the page; favicons check the page's icon first.
VoiceOver
- Toolbar, tabs (with “Tab 2 of 5: Title selected”), address bar, find, zoom, and bookmarks all properly labeled with hints and live announcements.
- Hidden tabs are hidden from VoiceOver; find bar and zoom percent are spoken.
Build 26 — browser bug fixes
- Fixed the scroll direction being inverted — "scroll up" now actually scrolls up.
- Fixed stale element references after opening a new page.
- The share sheet now anchors to the correct tab's web view.
- localhost with a port number opens correctly instead of triggering a search.
- Cleaned up unused code in the browser engine.
Build 25 — the browser works with VoiceOver
- Browser tabs are now real buttons: switch tabs, close tabs, and hear every tab announced by title.
- Agent status changes in the browser panel are announced as they happen.
- Toolbar and panel controls are fully labeled, and the start-page search field is announced.
Build 24 — Safari-style polish, live answers in the browser, downloads
- The browser window is styled like Safari: a unified toolbar merged with the title bar, and a full-width tab bar with favicons.
- Agentic's answers now stream live inside the browser panel — question, work, and answer all in one window.
- Downloads work: clicking a download link shows a save dialog.
- The agent now knows when the browser window is open and keeps web tasks inside it.
- Share button for the current page, and tab favicons.
Build 23 — the browser feels like a real browser
- Chrome-style keyboard shortcuts: ⌘T new tab, ⌘W close tab, ⌘L address bar, ⌘F find in page, ⌘R reload, ⌘[ / ⌘] back and forward, ⌘+ / ⌘− / ⌘0 zoom.
- Find-in-page bar with match count and next/previous.
- The address bar now searches Google when you type words, and opens URLs when you type an address — with select-all on focus.
- VoiceOver: tabs are real buttons, agent status changes are announced, and every control is labeled.
- Agent actions in the browser no longer steal keyboard focus or reset the agent panel while you type.
Version 1.3
August 2026
Attach images, PDFs, and text or code files to your messages; a rebuilt screen controller that sees clickable elements instead of guessing at pixels; a clearer activity log; one-click in-app updates; and a configurable step limit for longer jobs.
Build 20 — questions with tappable answers, no response-length cap, unlimited steps
- When Agentic needs information or a choice, it now asks with tappable suggested answers — and you can still type your own.
- The response-length limit is gone: the model decides how long its answers are.
- New "Keep going until the task is done" setting lets the agent work past the step limit (the stuck-action safety stop still applies).
Build 18 — crash fix for letter typing, unmissable update alerts
- Fixed a crash that could quit the app when it pressed or typed letter keys on screen (the keyboard-layout lookup ran off the main thread).
- The update dialog now appears whenever an update is found — including the automatic check at launch — not just for manual checks.
- Rapid reconnect clicks on tool servers can no longer create duplicate connections.
Build 17 — the app reopens itself after updating
- Fixed the app not reopening automatically after a one-click update (LaunchServices race right after the old instance exits — now retried up to three times with a forced fresh launch).
Build 16 — reliability fixes
- MCP tool servers now shut down cleanly when Agentic quits (local server processes no longer linger).
- A failed update no longer leaves its download behind, and update checks retry once on a flaky connection.
Build 15 — self-update made bulletproof
- The update is now installed while the app is still running, before it quits — a failed helper can no longer close Agentic without the update landing.
- The relaunch helper no longer waits on a recycled process ID (the cause of updates stalling forever), with a safety timeout as a backstop.
Build 14 — conversation summaries, VoiceOver polish
- Conversations are now summarized automatically: after your first exchange, the sidebar and menu show a short generated title instead of the first line of text.
- VoiceOver no longer reads the app name twice in the header, and spoken text now says "Agentic" everywhere instead of "the assistant".
Build 13 — reliable self-update, unmissable update alerts
- Fixed the one-click update closing the app without installing or reopening — updates now swap in and relaunch reliably.
- Check for Updates now shows its result in a clear dialog with Update Now / Download in Browser / Later buttons.
- Downloading in the browser is now only a fallback — the in-app update is the primary path.
Build 12 — single sounds, cleaner Settings
- Fixed response and error sounds playing twice (an overlapping system notification sound).
- Removed a redundant line from Settings.
Build 11 — new sounds, calmer stopping
- Fresh response, error, and cancelled sounds.
- Pressing Stop is no longer shown as an error, and Escape no longer cancels — only the Stop button does.
Build 10 — faster attachments, vision-aware sending, one-click updates, version display
- Attaching is instant, and images are downscaled before sending so vision answers arrive much faster.
- Images require a vision-capable model (the app tells you when to switch) instead of quietly converting them to text; PDF and text attachments keep working with every model.
- The update banner downloads, verifies, and installs the update itself, restarting Agentic automatically.
- Settings shows the exact version and build you are running.
Build 9 — screen control rebuilt, clearer activity log
- capture_screen lists every clickable element with exact coordinates from the macOS Accessibility API, and new
ax_click / ax_type tools press and fill elements directly instead of guessing at pixels.
- Multi-display capture fixed, and keyboard keys follow your keyboard layout.
- The activity log shows exactly what each step will do; the message bubble keeps only the final answer.
Version 1.2
August 21, 2026
- Custom AI provider. Connect any OpenAI-compatible server — one-click presets for GitHub Models (free tier), LM Studio, NVIDIA NIM, DeepSeek, and Mistral, or paste your own base URL. Model lists load from your server automatically.
- MCP tool servers. Add Model Context Protocol servers (a local command or remote URL) and their tools join Agentic's toolkit — always asking for approval before they run, with an "always allow" option.
- Longer tasks. The per-task step limit is now configurable in Settings (up to 200 steps) and the assistant stops itself if it notices it repeating the same action without progress — instead of getting stuck in loops on fiddly screen work.
Version 1.1
August 21, 2026
- Update checker. Agentic now checks for new versions automatically at launch, and you can check manually anytime from the Agentic → Check for Updates… menu. When an update is found, a banner appears in the app with a one-click download.
- OpenCode Zen provider. The OpenCode team's curated gateway, including a rotating set of free models fetched live in Settings.
- Live model catalogs. Settings lists every model each provider offers, with free vs. paid determined by actual pricing, not names.
- Markdown responses. Bold, code, and links render properly instead of raw symbols; the chevron logo is used everywhere in the app.
- Refined icon and branding, plus this What's new page.
Version 1.0
Initial beta · August 2026
- Five AI providers. A private local model (Ollama) by default — free and unlimited — plus Google Gemini, OpenRouter, Groq, and OpenAI with your own API keys.
- 32 built-in tools. Run shell commands, read/write/edit files, search the web and read pages, see and click your screen, type and press keys, manage the clipboard, set reminders, and do math.
- Subagents. Big jobs are delegated to focused researcher, implementer, verifier, or general subagents, then combined into one answer.
- Built for VoiceOver. Every step, approval, error, and finished response is announced out loud, with custom sounds so you know when something happens.
- Saved conversations. Every chat is saved automatically, listed in the sidebar and the menu-bar Conversations menu, ready to reopen anytime.
- You stay in control. Safe actions run automatically; anything destructive shows exactly what it will do and waits for your approval.