Say it to your phone.
Your assistant does it on your computer.
The assistant is the one agent of your account you can reach from anywhere: "send Marie the summary", "read my mail", "check my calendar", "handle the last bug of the shop project". No project to open, no agent to find. You talk, it acts, the answer lands as a notification.
It runs on your desktop through the agent CLI you already use, with your MCP connectors and the AgentsRoom tools that reach every project. The phone only records and sends, over the same encrypted link as the rest of the mobile app.
The phone records and sends; everything else happens on your computer, with your connectors.
Every assistant on a phone promises the same thing and stops at the same wall: it cannot reach your work. Your mail, your calendar, your messaging, your projects and the agents working on them live on your computer, behind your own accounts. A voice assistant that lives on a server somewhere sees none of it.
The AgentsRoom assistant turns that around. It is an ordinary AgentsRoom agent, running on your desktop through your own CLI, so it inherits what your other agents have: the MCP connectors you attached to your account, the backlog of every project, the ability to start a colleague on a ticket. What the phone adds is the shortest possible way to talk to it.
Siri, the iPhone Action button, a Back Tap, an Android widget or Quick Settings tile, the share sheet, the Assistant tab of the app: each of them opens the assistant's screen with the microphone already listening. You say one sentence, the transcript lands in the composer in case a name came out wrong, and you send it. You put the phone down. The answer is a notification.
A real session, filmed in the app: the personal assistant is told once about a half marathon, with the date, the goal and the coach. It files a "Running" dossier and a contact card in its private memory on its own, and in a brand-new conversation it still knows the race date and which coach to call.
Why one assistant for the whole account
AgentsRoom already lets you dictate a prompt to the agent you have open, and share a screenshot into its composer. That is the right tool when you are in the middle of a task. It is the wrong tool for "remind the team about tomorrow" from a car park, because it starts with a project and an agent you have to pick.
The assistant has no project to pick. It is one agent, global to your account, that knows where things are: it lists your projects, reads and files backlog tickets in any of them, resumes or launches an agent there, and reports back in a sentence. Its console runs in a folder of its own, never inside a repository, so no project's instructions ever leak into the conversation.
It also has a memory that outlives the conversation. Every day starts a fresh one, so the context never fills up with last week; if you allow it, what matters (a subject you follow, a contact, a rule you gave it) goes into a private memory of its own, read back when it is needed.
Reached in one gesture
Every entry point opens the same link, agentsroom://bot, on the assistant's screen with the microphone already listening.
Siri and the Shortcuts app
"Talk to Bot in AgentsRoom" is an App Shortcut: it exists as soon as the app is installed, with no setup. It also appears in Spotlight and in the Shortcuts app, where any automation can call it.
Action button, Back Tap, app icon
On an iPhone with an Action button, assign the "Talk to Bot" shortcut to it: one press, the assistant is listening. Back Tap (two or three taps on the back of the phone) takes the same shortcut. A long press on the app icon offers two quick actions, "Talk to the assistant" and "Write to the assistant".
Widget, tile and app icon
On Android, a 1x1 microphone widget on the home screen, a "Personal assistant" tile in Quick Settings, and the same two quick actions under a long press on the app icon. All of them open the assistant's screen directly.
The share sheet and the Assistant tab
Share a text or a link from any app to AgentsRoom and pick "Send to Personal assistant": an article to summarise, a message to answer, an address to put in the calendar. Inside the app, the Assistant tab sits in the middle of the tab bar and opens the same screen, to talk or to type.
And from the desktop
The Assistant button sits in the title bar of the desktop app, one click from every screen: its panel opens over whatever you were doing, as a conversation, with the raw terminal one toggle away and the same Voice Mode as any other agent. The phone is the fastest way in, not the only one.
What it can do once you have its ear
Whatever your other agents can do, plus the whole account.
Every project, every ticket, every agent
"Handle the last bug of the shop project": the assistant lists your projects, reads that backlog, and starts an agent on the ticket, or files a new one. It says in which project it acted.
It remembers you
A fresh conversation every day, and, if you turn it on, a private memory in between: one dossier per subject, your contacts, the rules you gave it. Nothing is re-explained the next morning.
It confirms before it sends
Reading, searching, summarising and drafting need no confirmation. Sending a message, deleting, publishing or paying does: the assistant shows the exact text and waits for a yes.
The answer comes to you
When the assistant finishes, or when it has a question, the phone gets a push with the summary. Tap it to open the conversation, or have the answer read aloud.
End-to-end encrypted, no server queue
What you say travels through the same encrypted relay as the rest of the mobile app. AgentsRoom's servers never see the text: there is no queue holding it in clear.
Desktop away? The phone says so
If your computer is unreachable, the phone says so before you speak, with a retry. A text shared from another app is kept on the phone, with a notification saying so, and goes the moment the desktop is back.
A memory of its own
If you want it to, the assistant keeps what matters to you, subject by subject, in a private memory your project agents never see.
One dossier per subject
Your building's management board, your investments, a running plan: each subject gets its own dossier, with the context, your goals and what is still pending. Ask about it next month and it picks up where you left off.
The people you deal with
One card per person: who they are, how to reach them, what you last discussed. "Tell Marie" needs no explanation the second time.
A dated journal of your requests
Each dossier keeps a journal of what you asked and what was done. The date is set by AgentsRoom, not guessed by the model, so "what did I send the board last month?" has a real answer.
Versions to undo
Every note keeps its previous versions. If the assistant got something wrong, restore an earlier one in a click, or undo right away from the line that says it kept something.
Photos kept with the notes
Share a photo from your phone or drop it on the assistant's panel: a recipe, a scanned letter, a training plan. The assistant files it in the right dossier, next to the note it belongs with, and shows it back when you ask. How much it can keep depends on your plan.
You choose what it keeps
The first time, the assistant asks. Your answer can be changed at any time in its settings, on the desktop and on the phone.
Nothing
Each conversation starts from scratch.
Ask me first
It suggests what to keep and saves only once you agree.
Automatic
It keeps what is useful on its own and tells you in one line.
Export the whole memory as Markdown (with its images, in a zip file) or JSON whenever you like, or erase everything in one click.
Where it is kept
- Encrypted on the AgentsRoom servers with a key specific to your account, so you find the same memory on every computer and on your phone.
- Only your assistant reads it. Your project agents never do, and it never ends up in a repository.
- Never stored: passwords, tokens, card numbers. A note that contains one is refused; keep those in the secrets vault.
- When the assistant reads a note, that note is sent to your model provider, like the rest of the conversation.
- Photos are compressed and cleared of their location data before being encrypted like the notes. When the assistant looks at one, it is sent to your model provider.
Specialists you talk to directly
Give your assistant a running coach, a wealth advisor, a home manager: each one is a tab next to it, with its own rules and its own dossier.
A real session, filmed in the app: one click on New specialist, a Running coach named with its purpose and standing instructions written once. It gets its own tab next to the assistant, answers a 12 km run report the way it was told (short, three runs a week at most) and keeps its week plan in its own dossier of the assistant's private memory.
One tab per specialist
Each specialist is a conversation of its own in the assistant panel, on the desktop and on the phone. Ask the coach about Sunday's long run without scrolling past your tax questions. With many of them, a search field finds the right one.
Standing instructions, like a CLAUDE.md
Tone, rules, what it must always do after a session: write them once. They apply from the first message of every conversation, and the specialist updates them itself when you give it a new rule. Stored encrypted with your account, with the rest of its memory.
Bring the memory you already built
Kept your coach in a git repository with a CLAUDE.md that commits after every change? Import the folder: its Markdown notes go into the specialist's dossier, its CLAUDE.md or AGENTS.md becomes its instructions. No repository to maintain any more, the same memory on every device.
A webhook can wake it up
A trigger can hand its prompt to your assistant or to one specialist instead of starting a new agent: the sync job that imports your runs tells the coach, and the analysis lands in that conversation, answered with its memory.
They hand tasks to each other
Ask your assistant about Sunday's long run and it offers to pass it to your running coach, whose field it is. Say yes and the brief lands in the coach's tab, and the answer comes back to the assistant. It can also create a new specialist once you agree.
One team, one view of your life
Each specialist reads what the others know before advising: the running plan shows a key session on Thursday, so you get a nudge to skip drinks and sleep early on Wednesday; a tight budget becomes one more reason. Reading the shared memory comes first, and messages between them are capped.
How it works, in practice
From the sentence to the notification.
Wake it up
Siri, the Action button, the widget, the tile, the share sheet, or the Assistant tab in the middle of the tab bar. The app opens the assistant's screen with the microphone already on.
Say it
One sentence. The AgentsRoom speech-to-text transcribes it, the same engine as the composer dictation: the quality is the app's, not the system dictation's.
Check, then send
The transcript lands in the composer. Fix a word if a name came out wrong, or send it as is. Prefer typing? "Write to the assistant" opens the same screen with the keyboard.
It reaches your desktop
The text travels over the encrypted relay. The desktop opens the assistant's console if it was closed, resumes today's conversation, and hands it the sentence.
The assistant acts
Through your CLI, your connectors and the AgentsRoom tools: it reads, files, launches, drafts. Anything visible to someone else waits for your yes.
The answer lands
A push notification with the summary. Open it to read the whole answer or to reply by voice, and if the task was long, the desktop calls you by your first name when it is done.
One page, on the desktop and on the phone
Settings > Assistant holds everything about it, and the phone shows the same rows with the same words: an edit on one side is the value on the other.
Name and avatar
"Personal assistant" is its default name, in the language of the interface. Rename it, give it a face: the title bar, the phone and the notifications follow.
Provider and model
Any agent CLI you have installed, with the same picker as every other agent. By default it follows the provider your quick agents use. For Codex, choose an assistant-specific Standard, Fast or Ultrafast service tier, or inherit your existing settings. The next-launch tier is shown; Ultrafast requires support from your model and account.
Launching agents without asking
Off by default: the assistant asks on the desktop before starting an agent in another project. Turn it on and "handle that bug" just goes.
Enable or disable
One switch. Off, the button leaves the title bar, the Assistant tab leaves the phone, and messages from the phone are refused.
New conversation
The assistant starts a fresh conversation every day; its memory carries what matters over. One button starts over right now.
Called by your first name
When a task longer than a minute ends, the desktop says your first name and what was done, in the Voice Mode voice. Off in one tap.
The assistant ships without a cloud fallback: your desktop has to be running (see the limits below).
What it does not do yet
Said plainly, so nobody discovers it in a car park.
- Your desktop has to be running and signed in. There is no cloud execution when the computer is off: your connectors and your CLIs live there, not on our servers. The phone tells you before you speak.
- The transcription is the AgentsRoom speech-to-text, the same as the composer dictation, billed on the same allowance and credits. Siri only opens the app; it never dictates for it.
- The assistant lives on one paired desktop at a time. With several computers, the phone sends to the one it is paired with, and offers the switch when another is online.
- On iOS the Siri phrase is fixed, in English ("Talk to Bot in AgentsRoom"), even after you rename the assistant. The Action button, Back Tap and, from iOS 18, the Talk to Bot control in Control Center or on the lock screen open the assistant without Siri.
Questions about the assistant
What is the AgentsRoom assistant?
The personal assistant of your AgentsRoom account: one AI agent you can reach from your phone in a single sentence, without opening a project or picking an agent. It runs on your desktop through your own agent CLI, with your MCP connectors and the AgentsRoom tools that reach every project, and answers by push notification. It is called "Personal assistant" by default; you can rename it.
Does Siri do the transcription?
No. Siri, the Action button, the widget and the shortcuts only open the AgentsRoom app on the assistant's screen. The app's own microphone records, and the AgentsRoom speech-to-text transcribes, the same engine as the composer dictation. That is what keeps the quality consistent across languages.
Where does it run?
On your desktop, in the AgentsRoom app, through the agent CLI you chose for it (Claude, Codex or any other installed one). Its console runs in a folder of its own under the AgentsRoom home directory, never inside a repository, and it is tied to none of your projects: you name the project a request is about, or it asks.
Can it read my mail or send a message?
If you attached the matching MCP connector to your account, yes: the assistant reaches it like any other agent. Reading and drafting need no confirmation; sending, deleting or paying does, and the assistant shows the exact text before it goes.
What happens if my computer is off?
The phone tells you right away, before you speak, and offers a retry. A text shared from another app is kept on the phone and sent as soon as the desktop is back. There is no cloud execution in this version: your connectors and CLIs are on your machine.
Is what I say encrypted?
Yes. The text travels through the AgentsRoom relay, which is end-to-end encrypted between your phone and your desktop. There is no server-side queue: a message the desktop cannot take stays on the phone.
Does it remember me?
The assistant starts a fresh conversation every day, so it never drags last week's context around. If you turn its memory on, what matters is kept in a private memory of its own: one dossier per subject, your contacts, a dated journal of your requests. You choose between Nothing, Ask me first and Automatic, and you can export or erase everything at any time.
Can I rename it or change its avatar?
Yes, in Settings > Assistant on the desktop or in the assistant card of the phone's settings. The name and the avatar follow everywhere: the title bar, the phone, the notifications.
Can it launch agents on its own?
Only if you allow it. By default the desktop asks you before the assistant starts an agent in another project. The setting "Can launch agents without confirmation" removes that step.
Which agent CLIs does it work with?
Any CLI AgentsRoom supports, picked in Settings > Assistant. A CLI with no MCP client answers but cannot reach your connectors or your other projects; the settings page says so when you pick one.
Who can read the assistant's memory?
Only your assistant. The memory is encrypted on the AgentsRoom servers with a key specific to your account, which is what lets you find it on every computer and on your phone. Your project agents never read it, and it refuses passwords, tokens and card numbers. When the assistant reads a note, that note is sent to your model provider like the rest of the conversation.
Goes well with
Mobile Companion
Dictate into the agent you have open, and share files into its composer.
Voice Mode
A spoken conversation with any agent, on the desktop.
Mobile and desktop sync
The whole workspace on your phone, in real time.
Project Memory
The memory your project agents share. The assistant keeps its own, apart.
Agent notifications
The push that brings the assistant's answer back to your phone.
AgentsRoom MCP
The tools the assistant uses to reach every project of your account.
One sentence away.
Download AgentsRoom and say "Talk to Bot" to your phone.
Companion app: monitor your agents on the go
Bring your own: Claude, Codex, Antigravity CLI, or other AI provider.
Push bugs and requests straight to your public backlog.