Changelog

What's new in Agents At Work.

October 11, 2026

Made for iPhone Duo

iPhone and iPad 2.9 (63) · Android 2.9 (37) now on Google Play for everyone

  • Open it in landscape and your sessions sit beside the feed. The side panel lists the computer's sessions, the feed takes the rest, and the back and menu buttons move to the bar by the camera.
  • Fold it halfway and the bottom half is your desk. The feed stays on the upper half; the lower half becomes a page for your prompt with the full keyboard under it, typewriter style, and it keeps your draft and focus when the keyboard comes and goes.
  • Answer on the closed phone. An agent's question shows on the outer display with its options, so one tap answers it without opening the phone.
  • Small things set right. Timestamps never run ahead of the clock (a computer a second ahead used to print "in 0s"), and the working indicator steps aside while a question is waiting for you.
  • Android 2.9 is out of open testing. Everything from October 7, image attachments included, is now in production on Google Play.

October 7, 2026

Send a photo to your agent

iPhone and iPad 2.8 (62) · Android 2.9 (37) · macOS 1.7.3 (75) · Linux: aaw-core 0.1.6

  • Attach an image to a prompt. Pick a photo from the session menu and it travels end-to-end encrypted to your computer, where the agent reads it. Works with every agent: paste a screenshot of a bug, a whiteboard, a design, and ask away. Nothing is stored on the relay.
  • Pair over SSH, no camera needed. aaw link now also prints a tappable link, so a Linux box can be paired from an SSH app on the phone itself.
  • The sidebar follows you into a session. On iPad and in landscape, the side panel shows the open computer's sessions so you can hop between them; Android gets a matching two-pane layout on tablets and foldables.
  • Know when your computer is running out of memory. One push warns you when the machine is critically low (sessions may freeze), a badge shows on the computers list, and it clears on recovery.
  • A prompt sent to a stopped session is never silently lost. It is delivered once on the next start, or, past 30 minutes, the feed says it expired so you can resend.
  • Payment hiccups keep access. Android now matches iOS: if a renewal fails, you keep working through the store's grace period with a banner pointing at the fix.

October 5, 2026

Browse a session's files, and no more zombie questions

iPhone and iPad 2.7 (61) · Android 2.8 (36) · macOS 1.7.2 (74) · Linux: aaw-core 0.1.5

  • Browse files from the session. The session menu now has Browse Files: it opens the folder browser right in the session's own folder, so you can check what the agent wrote and download a file without leaving the conversation.
  • An unanswered question can't freeze the session. If a permission question goes unanswered for 5 minutes, the computer now approves it, says so in the feed, and moves on; before, the dead question card could linger on the phone and hold up the next question for another 5 minutes.
  • Expired questions look expired. A question card past its deadline greys out and explains what happened instead of accepting taps nobody is waiting on, and its countdown now shows the real time the computer waits. Update the Mac app (or re-run the Linux install) so prompts shown in the terminal keep taking late answers.
  • Know when notifications can't reach you. If notifications are off, a banner warns that approval requests won't alert you and takes you straight to the right settings screen; Android also detects Do Not Disturb.

October 1, 2026

See the work that keeps going after the reply

iPhone and iPad 2.6 (60) · Android 2.7 (35)

  • Background agents, visible. When Claude answers but its background agents are still working, the session now shows "Waiting for N background agents to finish" until the results land, so a quiet feed no longer looks done too early.
  • A refused prompt tells you why. If the agent cannot run your prompt at all (out of credits, a usage limit), the reason now arrives in the feed within seconds instead of silence.
  • Know when your phone is the offline one. Airplane mode or no signal now shows a clear banner under the title bar; before, the app just looked quiet and the stale feed was a mystery.
  • Deleting asks first. Swiping away a session now confirms before stopping it on the computer, and removing a computer confirms before unpairing (re-pair anytime by scanning its QR code).

September 30, 2026

Restart without losing the conversation

macOS 1.7.1 (73) · Linux: aaw-core 0.1.3

  • Restart keeps your history. Restarting a Claude or Scoot session from the phone resumes the conversation on the computer, and the phone's feed now survives with it instead of going blank. A fresh conversation still starts with a clean feed.
  • One prompt, one entry. A prompt sent from the phone was stored twice behind the scenes; it now lands in the feed exactly once, while a genuinely repeated message stays repeated.
  • Self-hosting the relay? Update to aaw-core 0.1.3: device tokens are now stored hashed, the store migrates itself with a rollback snapshot, and the handshake carries a protocol version for future upgrades. Our hosted relay already runs it.

September 29, 2026

Plans made clear, and a working indicator you can trust

iPhone and iPad 2.5 (59) · Android 2.6 (34) · macOS 1.7 (72) · Linux: aaw-core 0.1.2

  • ☕ Plans, side by side. The plans screen shows Free, Personal and Pro together and opens on the plan you are on. Free is a real choice: one computer, no account, nothing to buy. Open it anytime from Settings, See Plans. Plans →
  • Upgrade or switch in the app. Your current plan shows Manage Subscription; the others show Upgrade or Switch. On Android, upgrading now replaces your plan instead of adding a second one.
  • The working indicator stays on until the agent finishes. Claude Code 2.x changed what its screen looks like mid-turn, and the indicator could vanish seconds into a long answer. It now holds for the whole turn. Update the Mac app, or run the Linux install command again.
  • The reader follows your phone's appearance. Long responses open in light or dark to match, on iPhone, iPad and Android.
  • Unlink asks first. Unlinking a phone now says how many computers it will forget before it does.
  • Smaller fixes. Session titles no longer go blank or get cut short on iPhone, read-aloud no longer slows the app down, and subagent reports no longer show up in the feed as if you had typed them.

September 29, 2026

Self-hosting the relay? Update to aaw-core 0.1.1

aaw-core 0.1.1 · relay only

  • Limits on every client. The relay now enforces frame size and rate limits, history size caps, per-computer storage and push budgets, and stricter rules for re-registering a computer.
  • Our hosted relay already runs it. If you run your own relay, update aaw-core to 0.1.1 and restart it. Computers and phones need no change; everything else in 0.1.1 is unchanged.

September 28, 2026

🎉 The engine is open source

aaw-core 0.1.0 · MIT license · macOS 1.6 (71) · Linux

  • 🔓 Everything that runs on your computer is open. The daemon that watches your agents in tmux, the hooks that report what they do, and the aaw command are now aaw-core, MIT licensed, on GitHub and PyPI. You can read every line of what runs next to your code.
  • 🛰 The relay is open too. The small server that carries messages between your computer and your phone is in the same repository. It only ever sees encrypted messages and where to deliver them, and you can run your own with docker compose.
  • 🍎 The Mac app runs it. macOS 1.6 bundles the same engine, so the app and a headless Mac or Linux install behave the same way.
  • 🐧 Linux installs it from PyPI. The one-line install command sets up aaw-core in its own virtualenv; running it again, or aaw update, keeps it current.
  • 🤝 Contributions welcome. Found a bug or want support for another agent? Open an issue or a pull request on GitHub. Scoot 🛴, our own agent, is open source as well, on GitHub.

September 28, 2026

iPad support, no accounts, and one computer free

iPhone and iPad 2.4 (58) · Android 2.5 (33) · macOS 1.6 (71) · Linux: aaw-core 0.1.0

  • A proper iPad layout. Your computers sit in a sidebar and the selected computer's sessions and feed fill the rest of the screen, in portrait and landscape. With a hardware keyboard, Return sends the message and Shift-Return starts a new line. A long response opens in a full-screen reader.
  • No accounts. The Mac app, Linux, and the phone apps no longer have a sign-in. You pair a phone with a computer by scanning a QR code, and that same scan sets up the end-to-end encryption. Messages travel through our relay, which only ever sees ciphertext; you can also run your own.
  • One computer is free. Following one computer on our relay costs nothing. The Personal plan raises that to 3 computers and Pro to 15, and an App Store subscription now covers every iPhone and iPad on the same Apple ID. Plans →
  • A voice you choose. Settings, then Voice, picks the voice that reads responses aloud, from the voices installed on your phone, with a short sample of each.
  • Moving from the earlier version. Update the phone app, and remove the old computer from its list (or Settings, Unlink this Phone). On the Mac, install the new app over the old one and click Link Mobile. On Linux, run the install command again or aaw update, then scan the new QR code. Sessions started under the old version do not carry over; start them again.

September 15, 2026

Linux: a scrollable, searchable feed and friendlier schedule times

Linux 0.2.0 (13)

  • Scroll and search the feed. An agent runs in a tmux screen with no scrollback, so a long conversation was stuck on one screen. aaw feed my-app now opens in a pager: scroll with the arrow keys, and search with / like in less. aaw feed my-app -f still follows the live feed.
  • The full history from the cloud. aaw feed my-app --remote pulls the whole retained conversation from the cloud, decrypted on your machine, the same source the phone and Mac read, instead of just what has happened since the session started.
  • 12-hour or 24-hour schedule times. aaw schedule now takes either, so --at 10pm and --at 22:00 both mean 10 PM. aaw schedule --help shows the accepted formats.

September 14, 2026

Schedule a prompt for later

macOS 1.5 (70) · Linux 0.2.0 (11) · iPhone and Android rolling out to the App Store and Google Play

  • Queue a prompt now, run it later. Set a prompt to run at a time you pick, e.g. 2 AM when your token limit resets, so a spent quota turns into work done overnight. The prompt waits until its time and then runs on the session you left going, on the computer's own clock.
  • From the phone, the Mac, or the command line. On iPhone and Android, open the session menu and tap Schedule a Prompt…, pick a time or a preset. On the Mac, use the schedule button in the Session Feed window. On Linux, aaw schedule my-app "run the tests" --at 2:10.
  • A queue you can manage. Every session shows how many prompts are waiting, with a calendar-clock marker on its row. Open the queue to edit the time or text of a scheduled prompt, or cancel it, from any device.
  • Private end to end. A scheduled prompt is encrypted on your device the same way messages are, and only your computer decrypts it when the time comes.

September 13, 2026

The right agent, and an easier-to-read feed

macOS 1.5 (69) · Linux 0.2.0 (10)

  • Fixed a session showing the wrong agent. If you started Scoot on a folder that had run Claude before, the session could show a Claude badge on your Mac and phone even though Scoot was actually running. The Mac app and the Linux background service now read the agent from the live session, so the correct agent shows everywhere.
  • Larger text in the Session Feed on Mac. The conversation is bigger and easier to read, with your prompts and the agent's replies at a consistent size.
  • A cleaner aaw feed on Linux. The terminal conversation view now wraps to your window and separates each turn, with a touch of color, so a long session is easy to scroll and read. Zoom the text with your terminal's own controls.

September 12, 2026

Open and save files from your phone

iOS 2.2 (54) · Android 1.1.1 (31) · macOS 1.5 (68)

  • View a file in the app. In the folder browser, tap a file to open it. Markdown is rendered formatted, and text and CSV files open in a clean, readable view, the same one used for agent replies. Works for .md, .txt, and .csv.
  • Save it to your phone. Download any supported file to your phone, to Downloads on Android and to Files on iPhone, straight from the viewer or the file's menu.
  • Tap a file for options. Files now open a short menu, View, Copy path, and Save, and the supported types stand out in the list.
  • Private end to end. File contents are encrypted on your computer and decrypted only on your phone, sent in chunks over the same secure channel as everything else, never stored on a server.
  • Polish: when a folder or file can't be read on a Mac, the app now points you to the exact System Settings screen to grant access; and the Start button on a stopped session no longer wraps on smaller phones.

September 11, 2026

Linux: a clear warning when run as root

Linux 0.2.0 (9)

  • Don't run aaw as root. Under root (or sudo) its state, keys, and the phone's folder browser point at root's home instead of yours, and updates can't restart the background service, so the phone would hang on the folder browser. aaw now warns loudly when started as root so this fails visibly instead of silently. Run it as your normal user.

September 10, 2026

Start a new session right from your phone

iOS 2.1 (53) · Android 1.1.0 (30) · macOS 1.4 (67) · Linux 0.2.0 (8)

  • Browse and launch. Open a computer, tap the new-session button, and browse its folders in the app. Pick a folder and start an agent there, no need to be at the Mac. Files are shown alongside folders for context, and repositories are marked.
  • Choose the agent, and the model. Pick Claude Code, Codex, Gemini, Grok, Cursor, or Scoot. For Scoot you also choose the provider and model. Tap Start and the session opens on your computer.
  • Copy a path. Press and hold any file or folder to copy its full path, ready to paste into an agent prompt.
  • Already running? If a different agent is already working in that folder, the app asks whether to open the new one alongside it or jump to the one that is already running.
  • Refreshed subscription screens in the Agents At Work colour, with Scoot listed among the supported agents.
  • Optional desktop notifications on Linux. Get a "waiting for your answer" alert on a Linux desktop when a question sits unanswered, set up in ~/.agent-bridge/config.json.
  • Polish: a larger send button in the session view on iPhone, and a cleaner "already running" dialog on Android.

September 9, 2026

Meet Scoot — local and cloud models in one agent

  • Scoot, a new agent. Scoot is our own open-source CLI agent (MIT, on GitHub). One tool runs free local models with Ollama and cloud models like OpenAI and Anthropic when you add your own key, with a router that can send a job to a cheap local model or an expensive cloud one. Install it with pipx install scootcli and Agents At Work detects it like the other agents.
  • Pick the model in the app. The Start Session window has a Scoot tab with a provider and model picker, and the session row always shows the model you chose, on the Mac and on your phone.
  • Ollama, now through Scoot. The standalone Ollama option is retired: the same free, private local models, reached through Scoot, plus cloud models and sessions. Any existing Ollama session still appears in the list.
  • The model shows in the session row wherever it is known, as "model · status".
  • Refreshed first-run walkthrough with the Agents At Work colour and larger, clearer text.
  • Fixes: phone approvals now reach a Scoot installed through conda or pyenv, and a slower phone answer no longer times out mid-approval; the app also backs up a hooks/settings file it cannot parse instead of overwriting it.

September 6, 2026

Agents At Work on Linux

  • Linux support, alpha. The same bridge to your phone, now for Ubuntu, Debian, Fedora, RHEL, and any other systemd distribution. There is no app: a background service does what the Mac app does, and one command, aaw, covers sign-in, pairing, and sessions. Install with curl -fsSL https://agentsatwork.app/install-linux.sh | bash.
  • Runs headless. Sessions stay bridged after you close the SSH window or log out, which makes a home server or a cloud VM a fine place to leave an agent working overnight.
  • Sign in from any browser. A headless box has no browser, so aaw login shows a one-time code that you enter on agentsatwork.app/link-linux from any device, then prints the QR code for the phone right in the terminal.
  • Plain claude, codex, gemini, grok, or cursor-agent in a project folder starts a bridged session, exactly as on the Mac. aaw start <folder> --agent <name> is the explicit form, and a second agent on the same folder runs alongside the first.
  • Mac and Linux side by side. A Linux machine shows up next to your Macs on the phone. Personal covers one computer, Pro up to five, in any mix.
  • Local models on Linux too. aaw start ~/proj --agent ollama --model qwen2.5 runs a session on a local Ollama model, no terminal involved, controlled from the phone like on the Mac.
  • Signed releases. The Linux installer and aaw update verify every release against the Agents At Work signing key before installing, on top of the checksum.
  • Also on the Mac: the terminal integration now covers cursor-agent, so a Cursor session started from a terminal is bridged like the others.

September 6, 2026

Sessions show "Ready" when they are ready

  • Fixed a finished session still showing as "Running" on your phone and in the Mac's session list until its next turn. The notifications that follow each reply were resetting the status a moment after it had been marked ready.
  • Fixed sessions failing to start after the Mac restarts. The first session after a reboot stopped at the keep-awake step and never launched the agent. Starting a second session worked, which is why it was easy to miss.
  • Checking an agent's version no longer starts a session. With the terminal integration enabled, codex --version, claude --help, or a script calling an agent in print mode could quietly start a full bridge session in the current folder. Only interactive launches start one now.

September 5, 2026

Messages from your phone always arrive, whole

  • Fixed a message from your phone occasionally never reaching the agent while the feed showed it as sent. The bridge now confirms delivery before showing your message as sent, retries on a hiccup, and tells you in the feed if it could not get through so you know to send it again.
  • Long and multi-line messages arrive intact. A long answer with several paragraphs could reach the agent as just its last line. Messages are now handed to the agent as one block, on Claude Code, Codex, Gemini CLI, Grok, and Cursor.
  • A message sent while the agent is showing you a question now waits until you have answered it, instead of being typed into the question.
  • Removing an Ollama session on the Mac now asks first, the same way stopping any other session does.
  • Fixed "restart" from your phone starting a fresh conversation instead of resuming, for projects whose folder path contains an underscore, a dot, or a space.
  • Agent sessions you start outside the app no longer show up in it. A Claude, Codex, or Gemini session launched from an IDE or a plain terminal could appear on your phone as a stray project, or write into a real project's feed.
  • Fixed push notifications silently not being sent on some Macs, depending on which Python the system had installed. The bridge now always uses its own bundled Python for them.
  • Fixed the Mac being kept awake after all sessions had stopped, when two sessions had been started at the same moment.
  • Fixed restarting a stopped session while a second agent was still running on the same folder: the restart could quietly land on the running agent instead, so the new session never appeared and messages went to the wrong agent.

September 2, 2026

Run a second agent on the same project

  • Open Codex alongside Claude on the same folder, or any other pair: one agent developing while another reviews, or a second agent picking up when the first hits its usage limit. Start Session with a different agent on a folder that already has one running, and the Mac asks whether to open it alongside or switch to the one already there. Each session gets its own terminal, its own feed, its own notifications, and its own Stop button.
  • Works from the terminal too. Typing codex in a folder where Claude is running now starts a separate Codex session. Previously it dropped you into Claude's window and confused the bridge about which agent was running.
  • Ollama can run next to another agent on the same folder. Before, starting Ollama on a folder with a running session quietly took over that session's connection to your phone.
  • The session list groups sessions by folder, with the agent named on each row when a folder has more than one.
  • Two folders with the same name in different places no longer collide; the second one is told apart by its parent folder.
  • On your phone, sessions on the same project sit together, titled by the project folder, each with its agent badge, so a project running Claude and Codex reads as one project with two agents rather than two unrelated entries.
  • Read-aloud skip now jumps to the next sentence instead of a fixed number of seconds, which kept landing mid-sentence.
  • Fixed a stale "stopping" spinner that could appear on a Mac session you had stopped and then started again.

August 30, 2026

Fewer ways to lose track of what your agent is doing

  • Tapping a notification on Android could close the app instead of opening the session it was about. It now opens the right session, and Back returns to that Mac's session list instead of jumping all the way out to your list of computers.
  • Your Mac now tells you when an agent is waiting and your phone is silent. If your phone is in Focus or Do Not Disturb, a push notification can be held back, so an agent could sit waiting for your approval without you noticing. The Mac app now shows a notification and flashes its menu bar icon if a question goes unanswered. Set the delay, or turn it off, from the menu (Alert If Waiting).
  • A heads-up when the connection to Claude is struggling. When the connection drops repeatedly, Claude retries on its own, but your phone showed nothing but "working…", which looked identical to a long task. You now get one notification when it's genuinely struggling, not one per retry. No action needed unless it keeps failing.
  • Messages sent to a stopped session no longer disappear. The message box on your phone used to let you type and send to a session that had already stopped, with nothing running to receive it. A stopped session now shows "Session stopped" with a Start Session button instead.
  • Stopping a session on the Mac now asks first, so a stray click can't end a running session. Also fixed a rare problem where the Mac app could freeze and start using a lot of CPU after stopping one.
  • Delete an Ollama session from your phone. This used to require going to your Mac. Long-press the session on iPhone, or use the ⋮ menu on Android.

August 29, 2026

Clearer messages when Claude's connection drops or your sign-in expires

  • A dropped connection, an expired sign-in, or a usage limit could previously show up in your feed as a raw, technical error message, or not reach your phone at all.
  • These now show a plain message explaining what happened and what to do next — for example, whether it's a brief connection issue you can just retry, or something that needs you to sign back in on your Mac.

August 29, 2026

Fixed a type of multiple-choice question that never reached your phone

  • Some of Claude's multiple-choice questions — the ones that show a preview next to each option — weren't recognized by the bridge at all, so they never reached your phone, even with mobile mode turned on.
  • These now show up and can be answered from your phone like any other question.

August 20, 2026

Formatted text in the expanded message view on iPhone

  • Tapping "Expand" on a long response now renders headings, lists, tables, code blocks, and links properly, matching what you already see on Android and in the regular conversation feed — previously it showed raw formatting characters instead.
  • Read-aloud now switches back to formatted text as soon as you pause, not just when you stop.

August 20, 2026

Fixed a session that could silently stop responding

  • In rare cases, the background bridge process for a session could get stuck internally and stop delivering messages, while still appearing to be running normally.
  • The bridge now watches for a message that isn't getting through and automatically restarts the affected session's connection within a couple of minutes, with no action needed on your end.

August 19, 2026

Fixed stopped sessions still showing as active on your phone

  • Closing a session on your Mac — or its terminal window closing any other way — wasn't always updating its status, so it could keep showing as active on your phone indefinitely, including a stuck "is working" indicator on iOS and Android.
  • Fixed a separate case where a reply could finish on your Mac but never reach your phone, after the bridge had been running for a long stretch without a restart.
  • The bridge now double-checks every session's status in the background and corrects it within a couple of minutes if anything gets out of sync.

August 16, 2026

Fixed a feed mix-up when switching sessions from a notification on iPhone

  • Fixed a bug where tapping a notification for one session while viewing a different session's feed could leave the old session's messages and badge on screen instead of switching to the new one.

August 15, 2026

Cursor support added

  • Claude Code, Codex, Gemini CLI, and Grok are now joined by a fifth agent: Cursor (Anysphere).
  • Cursor sessions show up in the mobile app with their own color-coded badge, the same push notifications and one-tap approvals as the other four agents.

August 15, 2026

Fixed messages from your phone silently failing to reach a long-running session

  • Fixed a bug where, after a session had been running for several days, messages sent from your phone could stop reaching your Mac — while replies from your agent kept arriving normally, making it easy to miss until nothing happened after you sent something.
  • The bridge now automatically refreshes its connection in the background on a regular schedule, and cleans up more reliably after a restart, so this shouldn't need a manual fix going forward.

August 9, 2026

Answered questions now stay in your conversation

  • Fixed a bug where, after you answered a question from your agent, both the question and your answer would disappear from the conversation with no record of what was asked or how you responded.
  • Fixed a bug where "Auto-approve" could silently pick the wrong answer on a question with more than a simple Yes/No choice. Auto-approve now only answers genuine Yes/No prompts automatically and still asks you for anything else.

August 9, 2026

Fixed a navigation bug on iPhone

  • Fixed a bug where opening a session from a notification, then tapping Back, took you to the list of computers instead of back to that computer's sessions.

August 9, 2026

Fixed a bug where a background task update could show up as a chat message

  • Fixed a bug where, after Claude finished a task running in the background, an internal status update could appear in your conversation feed as if it were a real message.

August 6, 2026

Fixed a bug where starting a new Claude session could fail silently

  • Fixed a bug where starting a Claude Code session for certain projects could exit immediately with no explanation, instead of starting normally.

August 6, 2026

Faster message sending and better formatting on Android

  • Sending a message now appears in the conversation instantly, matching the iPhone app.
  • Long agent responses opened full-screen now show proper formatting, headings, bold text, and code blocks, instead of plain text.

July 5, 2026

Session reliability fixes across Claude, Gemini, and Ollama

  • Fixed a bug where a new Ollama session's conversation window wouldn't update live until you closed and reopened it.
  • The session window now shows whether that session is actually running, with a one-tap restart if it's stopped — instead of looking active when it wasn't.
  • Fixed a bug where Claude could occasionally show an unrelated usage notice instead of its actual reply.
  • Fixed a bug where sending a message from the Mac's session window could unexpectedly switch the session into mobile mode.
  • Fixed a bug where Gemini's replies could take up to a couple of minutes to reach your phone or the session window.

July 5, 2026

Ollama session reliability fixes

  • Fixed a bug where an Ollama session could fail to start entirely due to a missing file in the app bundle.
  • Fixed a bug where a running Ollama session wouldn't show up in the Mac app's session list, even though it was reachable from your phone.
  • Fixed a bug where restarting an Ollama session could clear the conversation shown in the session window until it was reopened.

July 4, 2026

Fixed a startup bug affecting some fresh Mac installs

  • Fixed a bug where the Mac app could get stuck showing a plain yellow window at launch instead of the menu bar icon, with no way to sign in.
  • Fixed a bug where a fresh install could fail to start a session with a missing-dependency error.

June 29, 2026

Run local AI models from your phone, Ollama is here

  • Ollama is now a first-class agent in Agents At Work. Install Ollama on your Mac, pick a model, and control it from your iPhone or Android phone exactly like you would Claude, Codex, Gemini, or Grok, with live feed, push notifications, and one-tap replies.
  • Nothing leaves your desk. Your prompts go phone → your Mac → your local model → back to your phone. No third-party API, no per-query cost, no data sent to anyone's cloud. Ideal for private documents, sensitive projects, or anyone who just wants to keep things local.
  • Two modes depending on the model. Any Ollama model works in Chat mode for conversation. Models that support tool use (qwen2.5, qwen2.5-coder, qwen3) unlock Agentic mode; the model can read and write files, run shell commands, and complete multi-step tasks on your Mac while you're away from it.
  • Always asks before acting. In Agentic mode, the model pauses and asks your approval before running any destructive command, such as deleting a file, overwriting one, or running a shell script. You get the prompt on your phone as a tap-to-approve card, or inline in the session window on Mac. You can approve once or for the whole session.
  • Works while your Mac is asleep. As long as your Mac is plugged in and connected to the network, Ollama keeps responding to prompts from your phone even with the lid closed.
  • Start from your phone. If your Ollama session is stopped, you can restart it directly from the mobile app, no need to go back to the Mac.

June 26, 2026

Session feed window for Mac

  • A new window in the macOS app shows the live conversation with any active agent session, so you can follow what your agent is doing without keeping a terminal in view.
  • You can send prompts to your agent directly from the window using the text bar at the bottom (⌘↩ to send), without switching to the terminal.
  • The window includes a search bar to filter the conversation by keyword, and a clear button to wipe the display without affecting the session log on disk.
  • Session feed windows now appear in ⌘Tab while open and support ⌘` for cycling between multiple open sessions.

June 21, 2026

Restarting or clearing a session now actually clears your history

  • Typing /clear in the Claude Code terminal now deletes that conversation's history from our servers, not just your phone's screen — previously it stuck around in the cloud even though Claude had already forgotten it.
  • Sending /restart from your phone now clears it the same way, so the new session starts with a genuinely clean slate instead of stale history sitting in the cloud.

June 21, 2026

More prompts now reach your phone

  • Claude Code's plan-approval confirmation (the multi-option "Would you like to proceed?" menu) now reliably reaches your phone — previously it could fail to show up on mobile entirely.
  • Claude Code's interactive question prompts now show up on your phone as tappable question cards too, not just the standard approve/deny prompts.
  • Fixed Codex hooks silently breaking after a recent Codex update deprecated a config flag the bridge relied on.
  • Fixed a spurious "usage updated" notification appearing on your phone every time a session restarted on the Mac, even with no real usage change.

June 20, 2026

Agents At Work is public

  • The iPhone and Android apps are now publicly available, no more invite or beta enrollment needed. Download on the App Store or Google Play.
  • The macOS companion app remains a direct download, no app store required.

June 12, 2026

Grok support added

  • Claude Code, Codex, and Gemini CLI are now joined by a fourth agent: Grok (xAI).
  • Grok sessions show up in the mobile app with their own color-coded badge, the same push notifications and one-tap approvals as the other three agents.

May 30, 2026

End-to-end encryption

  • Conversation content, prompts, responses, tool-call summaries, and project paths are now encrypted with AES-256-GCM before they ever leave your Mac.
  • The encryption key is exchanged automatically during the QR scan that links your phone, no extra setup step.
  • Available on both iOS and Android.