Team AI for Mac — Release Notes
What's new in each release
Version 2.0.0 — A new look, safer tools, and the optional Team AI Plus
Everything you have today keeps working without an account. Team AI Plus is optional and nothing you already have is taken away.
New
- A redesigned Mac app — a quieter chat with unboxed replies and a model header, a model chip in the toolbar that opens the model picker, one overflow menu for the chat tools, and a glass composer. The system now draws the sidebar and toolbar.
- Team AI Plus (optional subscription) — Team AI for Mac is still a one-time purchase and works without an account. Plus is an optional subscription (monthly or yearly, paid on a Stripe checkout page in your browser; Team AI never sees your card) that adds encrypted sync across your devices, expiring encrypted share links, Team AI Included (replies without your own API key, within a monthly allowance; it is off until you choose it for a chat and agree to a notice, and your chat text then goes through Team AI’s server to Google’s Gemini API) and always-on agents (scheduled jobs that run on Team AI’s server; the prompt of a job is readable by the server and is not end-to-end encrypted). Plus needs a Team AI account: on the Mac, sign in with Google or a link emailed to you (Sign in with Apple is not offered on the Mac). Manage or cancel with Manage subscription in Settings → Team AI Plus. If Plus cannot be confirmed for 7 days the app goes back to Free; your chats are never affected.
- Charts and tables in replies, switchable per chat from the Tools menu.
- Better reading and chat controls, a hands-free voice mode, and backup, restore and import.
- Sharing your models on your network (off by default, encrypted, pairing by QR or link), and remote MCP servers with Connectors and sign-in.
Safer tools and Computer Use
- Tools now ask first in normal chats: anything that changes something (Computer Use actions, Shortcuts, MCP server tools, writing a file, creating a reminder or event) shows an approval window with what it is about to do. Your Always allow choices stay on this Mac.
- Run JavaScript shows the whole script and runs in its own helper process, stopped after 5 seconds or on excess memory, with no network or file access.
- The Computer Use emergency stop (Option-Command-Period) works, and a “Computer Use is active” indicator with a Stop button stays on screen.
- Local MCP servers need approval on each Mac, start correctly when Team AI is opened from Finder or the Dock, and keys are hidden in status messages.
Fixed
- The notarised app was missing the permissions to ask for the microphone, calendars/reminders and contacts, so dictation, Meetings, Live Voice and the calendar and contact tools could fail silently. Fixed.
- Many reliability and privacy fixes from the October 2026 audit, including deleted-chat crashes and a licence-trial protection fix.
Version 1.6.0 — Better image editing, a real Translator, and companion characters
Four additions, three of them requested directly by customers.
New
- Prompt-guided AI Fill — Remove Object's mask tool now takes an optional prompt: draw a mask and describe what should appear there instead of only erasing it. Leave the prompt blank for the original erase-and-fill behaviour.
- Remove Background — a new one-tap tool that lifts the subject out of a photo onto a transparent background, entirely on-device.
- Translator — a proper Translator screen for English, Turkish, German and Japanese, on-device via Apple's Translation framework (requires macOS 26 or later).
- Faster image generation — Fast/Balanced/Quality presets for on-device image generation, plus a live preview as the image forms instead of a blank progress bar.
- Companion Characters — assign one of 15 illustrated companions to any agent from a new picker, with a live rotating 3D preview. (AR Quick Look is an iPhone and iPad feature and is not in the Mac app.)
Version 1.5.0 — Ask your files, meetings, research, voice, Computer Use
Five major additions, all built around the same promise: your data stays on your Mac unless you choose otherwise.
Correction (3 Oct 2026): Your data is stored on your Mac, but Team AI also mirrors it to your own private iCloud while you are signed in to iCloud (on by default; the current version, 1.6.0, has no switch to turn it off), including the text of knowledge bases, watched folders and meeting transcripts. Anything you send to an AI provider you chose goes to that provider, and the Mac app contacts our licence server (licence key, a random device ID and your Mac's name) and our update feed. None of that includes your chats or files.
New
- Ask Your Files — Point Team AI at a folder and it keeps a private, on-device index of the documents inside: PDFs, Word files, Markdown, text, code and .eml messages. Changes are picked up automatically while the app is running; deleted files drop out. Attach a folder to any chat and every answer cites the file it came from — click to preview or reveal in Finder. Press ⌘⇧Space anywhere for "Ask my Mac".
- Correction (3 Oct 2026): The index is built on your Mac, but its text is saved with your other data and so follows the iCloud mirror, and passages used in an answer go to the AI provider you ask with unless it is an on-device model.
- Meetings & Voice Notes — Record from the microphone, the system audio of a call, or both. Transcription runs entirely on your Mac; you get a summary, decisions and action items that become Reminders with one click, and the transcript lands in a "Meetings" knowledge base you can chat against later.
- Deep Research — Ask a question and an agent plans sub-questions, reads sources and writes a report where every claim carries a numbered citation to a source it actually read; anything it could not verify is listed as such. Run it against the web or against your own files only. Export to PDF, re-run weekly as a Standing Agent, or run again and compare what changed.
- Live Voice — Talk to any agent hands-free and interrupt it mid-sentence. OpenAI Realtime or Gemini Live with your own key, or the fully on-device engine. Anything that would change data is read back to you first.
- Computer Use — Off by default. When you turn it on (Settings → Computer Use) and grant Accessibility, an Autonomous Agent can read the frontmost app's controls and click, type and choose menu items to complete a goal. Every action that changes something is shown to you first unless the app is on your allow-list; there is a per-run action budget, an ⌥⌘. kill switch, and password fields are never typed into.
- Shortcuts and widgets on the Mac — All twelve App Intents (Ask, Summarise, Translate, Save/Search Memory, and more) now work in Shortcuts and Spotlight on the Mac, and the Memory Digest and Standing Agent widgets are available in Notification Centre.
- 17 new local models — Qwen 3.5, 3.6 and 3.8, Gemma 4 QAT (2B to 31B), Liquid LFM2.5, IBM Granite 4.1, NVIDIA Nemotron 3.5, Meta Muse Glimmer and Cohere North Mini Code, on an updated on-device engine.
- Remote MCP servers over Streamable HTTP — the current MCP transport, with DeepWiki and Context7 in the directory ready to install.
Fixes & stability
- Standing Agents can now use read-only tools and can run when a watched folder changes.
- PDF export now paginates properly — long chats no longer lose everything after the first page.
- Current model lists for Anthropic and Gemini, and an up-to-date cost table in Usage & Costs.
- Four local models whose downloads had gone stale are back, with correct sizes shown.
- Your data now lives in an App Group container so the widgets can read it; the move happens once, on first launch, and keeps a copy of the previous store untouched.
Version 1.4.0 — Autonomous Task Agents
The biggest update since 1.1.0: a new agent type that acts, not just answers.
New
- Autonomous Task Agents — Give an agent a goal instead of a fixed prompt, and it plans and executes its own sequence of tool calls to get there, re-planning after every result. Any action that writes data or has a real-world effect pauses for your explicit approval first — you always see exactly what it's about to do before it happens. Generated images and spend are shown right in the run and tracked in Usage & Costs. Find it next to Standing Agents in the Agents tab.
- Search Everything — A new search (⌘F) across chats, messages, agents, memories, and Knowledge Base documents at once, instead of hunting through each section separately. Opening a message result jumps straight to it in the conversation.
- Standing Agents now tell you why — Agents show a "next run ≈" estimate, and if a scheduled run fails (a bad key, a provider error), you'll see why instead of just a stale "last ran" time.
Fixes & stability
- Hardened the streaming-response path against a rare crash class reported via TestFlight.
- Fixed a data race in model-list fetching that could occur when two screens looked up models at the same time.
- Fixed Ollama tool-calling: multi-tool-call turns now correctly correlate each result to the call it answers.
- Fixed Remove Object and Colorize to no longer crash on very large photos, and no longer return a smaller image than the one you started with.
- Fixed a download-cancel race that could leave a model download running after Cancel was tapped.
- Fixed a Knowledge Base document counter that could drift after a failed re-import.
- Global Search is now faster on large chat/document libraries, and debounced so typing doesn't stall the interface.
- A broad pass over the app's typography and spacing for more visual consistency across every screen, on both platforms.
Version 1.3.0 — Menu Bar Mini-Chat & Smarter Memory
A quick way to chat without opening the main window, plus more control over how Memory understands your conversations.
New
- Menu Bar Mini-Chat — a compact chat popover right from the menu bar icon, showing your most recent conversation with its own input box, so a quick question doesn't need the full app window open.
- Custom OpenAI-compatible embedding provider for Memory — point Memory's embedding step at any OpenAI-compatible endpoint you've configured in Settings › Providers (including a local Ollama server), with a model picker for both OpenAI and custom providers.
- On-device Ollama image generation provider in Image Studio.
Improvements
- Memory now resurfaces revisited conversations by actual recency, not just when they were first created.
Fixes
- Fixed memory extraction, the weekly digest, and Standing Agents silently failing to produce anything when using the on-device Apple Intelligence provider.
- Fixed the Menu Bar Mini-Chat sending messages into the wrong conversation when a group discussion or model comparison had been created more recently than your last single chat.
Version 1.2.0 — Device Limit, Now Enforced
Your license key is now checked against a real device limit (one Mac per license, as it has always been sold). Activating on a device already at the limit shows a clear message instead of a generic error, with a way to see and free up the device using your license from Settings.
Version 1.1.0 — On-Device Tools & a Refined Mac Experience
The first major update since launch, adding a full set of private, on-device tools and a redesigned, professional-grade Mac interface. It's still a one-time purchase — no subscription.
Highlights
- Knowledge Bases (Chat with Your Documents) — Add PDFs, text, and Markdown files and chat against them. Documents are embedded and searched entirely on-device, then attached to any chat from its tools menu. Your files never leave your Mac.
- Correction (3 Oct 2026): A knowledge base is embedded on your Mac by default, but its text and embeddings are saved with the rest of your data, which Team AI mirrors to your own private iCloud while you are signed in to iCloud (on by default; the current version, 1.6.0, has no switch to turn it off). A passage you attach to a chat is sent to the AI provider that chat uses (unless it is an on-device model), and if you choose a cloud embedding provider for a knowledge base its text goes to that provider. Team AI itself has no server that receives your files.
- Standing Agents — Set an agent to run on a recurring schedule and have its latest results waiting for you.
- Roundtable Listen mode — Hear a multi-agent discussion read aloud, with a distinct voice per agent.
- Weekly Memory Digest — A private, on-device "What I've learned about you" summary, refreshed about once a week.
- Correction (3 Oct 2026): In the current version (1.6.0) the digest is written by the first of your configured AI providers that answers, which can be a cloud provider (it then receives the memories it summarises). The digest is stored as a memory, so it is mirrored to your private iCloud like the rest of your data.
- One-Tap Boost — Have the model critique and refine its last answer in a single click.
- AI Judge for Model Comparison — Compare two models side by side and let an AI judge score which response was better.
- MCP Server Directory — Browse and add Model Context Protocol servers from a built-in, one-click directory.
- Decision Records — Key decisions from group discussions are saved as durable memories.
- Image Recipes — Apply chained image-editing steps to any generated image in one click.
- Monthly Spend Budget — Set a budget and track usage in Usage & Costs.
Design & Interface
- Redesigned three-column layout — a native, Mail-style source list → items → detail, with Settings now in the sidebar.
- Liquid Glass chat bubbles — chat and group-discussion bubbles now use Apple's Liquid Glass material, tinted per participant, in both light and dark.
- Full Light / Dark / System theme support — the whole window, all three columns and the title bar, switch together and stay consistent.
- A refreshed, more professional look throughout: a deliberate accent color, calmer empty states, and a refined splash screen.
Fixes
Data & sync
- Fixed an issue where updating from an earlier version with iCloud sync on could fail to open your data on launch, temporarily falling back to a blank in-memory session (your real data on disk was never at risk).
- Fixed data loss when regenerating a chat response, and a consent-screen bypass.
- Fixed data loss when interrupting or resuming a Group Discussion, and discussions occasionally assigning agents random playbook roles.
- Fixed a gap where a Workflow's pipeline progress could fail to persist.
Memory & Knowledge Bases
- Fixed Memory on Mac: the on-device summarization bridge was never actually wired up, plus deduplication and search bugs.
- Fixed Knowledge Base documents creating duplicate chunks on re-import, and a dangling attachment reference left behind after deleting a base.
- Fixed the Weekly Memory Digest silently failing to generate when its first-choice provider errored instead of trying a fallback.
Agents, Group Discussions & MCP
- Fixed Group Discussion voting cancellation, a turn-limit (maxTurns) bug, and per-agent tool assignment.
- Fixed the agent editor silently swapping a selected model, a duplicate model-list fetch, and a missing "Tools" section on Mac.
- Fixed MCP server environment variables (API keys, etc.) not reliably surviving save/reload, and the MCP Directory's "already installed" detection.
Interface
- Fixed the theme selection (System/Light/Dark) not fully applying to every pane, and made all three columns plus the title bar one consistent flat sheet in both themes.
- Fixed a Knowledge Base view getting "stuck" in the detail column after switching sections.
- Fixed chat auto-titles that could echo the conversation instead of summarizing it.
- Fixed malformed output when a sandboxed JavaScript tool hit a syntax error.
- Numerous smaller alignment, color, and stability fixes.
Known Limitations
- On-device (local) AI models require Apple Silicon. Cloud providers (bring-your-own-key) work on any supported Mac; the local MLX models do not run on Intel Macs.
- Standing Agents run when the app is open (macOS has no background scheduler equivalent to iOS); they fire on launch and on their schedule while the app is running.
- Knowledge Base ingestion is optimized for text-based PDFs, Markdown, and plain text.
Version 1.0.1
Initial public release of Team AI for Mac.