Docs / Concepts
How your Stand works
Summon your superpower. Sutando is your AI Stand — a personal spirit that fights on your behalf. It handles voice, calls, meetings, and your screen — and works while you sleep. This page is the two-minute mental model: the channels you summon it through, what stays on your Mac, and why messaging replies sometimes feel a beat slower than voice.
The big picture
Your agent lives inside the AG2 Space desktop app on your Mac. The app's rooms and DMs are its home channel; connect the other surfaces you already use and any time you talk to it through one of them, the same agent on your Mac handles the request.
Which channel for which kind of ask
Instant vs. background, in detail
Every request Sutando handles falls into one of two lanes:
Instant lane
Few hundred milliseconds
- Glancing at your screen and describing it
- Checking the current time / date
- Pressing a touchtone digit during a phone call
- Hanging up
- Recalling something from earlier in the same conversation
Things Sutando can answer in the same breath as your question — happens during a live voice or phone call without breaking the conversation.
Background lane
Seconds to a few minutes
- Drafting + sending an email
- Summarizing your unread inbox
- Booking a restaurant (placing a real phone call)
- Joining a Zoom meeting on your behalf
- Searching, reading, and quoting a webpage
- Posting to X, scheduling reminders, editing files
Things that need multiple steps — your agent hands these to its full-power core and pings you when the result is ready.
That's why messaging channels (Telegram, Discord) sometimes feel a beat slower than voice — they're nearly always the background lane. Voice and phone use the instant lane and can hand off to the background lane when needed.
Works while you sleep
About every five minutes, your agent wakes itself up and runs through a short checklist:
- Process anything in its queue (delegated tasks, messages received while idle)
- Check the health of your integrations and reconnect any that dropped
- Surface unanswered questions you'd asked it to remember
- Pick up the next item on whatever you're building together
Close your laptop or work on something else — tasks resume when the laptop wakes. If you don't want it pre-empting you, "presenter mode" pauses the background loop entirely.
A living system that learns with you
Your agent is not a fresh chatbot each time you open the app. It's a living system — the longer you work with it, the better it gets at being your assistant. Every conversation can leave behind a small note that future conversations will read, so corrections, preferences, and context accumulate over time instead of evaporating at the end of each session.
The notes it keeps are deliberately small — not entire transcripts, just the things worth carrying forward:
- Your profile (role, preferences, working style)
- Feedback you've given ("don't summarize at the end," "always cc this person")
- Project context (what you're building, deadlines, decisions)
- Pointers to external resources (Linear projects, Slack channels, dashboards)
Everything lives in plain markdown files on your Mac. You can read, edit, or wipe them at any time. Memory is re-read whenever a new request comes in, so changes take effect immediately — and the accumulated context gets richer with use, not noisier.
Permissions the app asks for
During first launch macOS will ask you to grant a handful of permissions. Each one unlocks a capability — skip the ones you don't want and everything else still works, just with fewer powers:
- Microphone — needed for voice mode
- Screen Recording — needed for "look at my screen"
- Accessibility — needed to click and type into Mac apps for you
- Contacts — needed to look up people by name before emailing or calling
- Calendar & Reminders — needed to read and create events / to-dos
- Notifications — so your agent can ping you when a background task finishes
Two engines, one assistant
Under the hood your Stand uses two AI providers, each chosen for the lane it's best at:
- Google Gemini (Live API) — powers the instant lane: voice, phone, multimodal (camera, screen).
- Anthropic Claude (via Claude Code) — powers the background lane: planning, tool use, multi-step jobs, code.
You can either pay each provider directly (bring your own keys) or let the managed plan handle billing for both. Either way, your data flow is identical — only the billing is different.
Where your data lives
The agent is local-first. Your conversations, notes, memories, and integration tokens live on your Mac as plain files — nothing is uploaded to a cloud server by default.
- Voice + phone audio goes to Google Gemini Live for real-time understanding, then is discarded after the response. Google doesn't save API audio.
- Background-task requests go to Anthropic Claude (via Claude Code), with the prompts + the tool calls your agent needs to fulfill them.
- Telegram and Discord messages traverse those providers' servers the same way any message you send through them does.
- Managed-plan tokens are minted by the cloud on the fly when you bring no keys; your audio and prompts still go direct to Google / Anthropic, not through Sutando's servers.
Sutando Cloud — for multiple devices
If you only use AG2 Space on one Mac, you can ignore the sync side entirely. It is opt-in and exists for one main reason: syncing across multiple devices.
- Run AG2 Space on your work Mac and your home Mac, and have memory, notes, and pending tasks follow you between them.
- Pick up a conversation thread on a phone-only access (no laptop open) using your account.
- Handle billing for the managed AI plan, so you don't maintain Gemini / Anthropic accounts directly.
You decide what to sync. The cloud never holds a copy of your local memory or transcripts unless you turn that on.
That's the whole model
You talk to your agent through whichever channel is closest to hand. The work happens on your Mac. The cloud only sees what you explicitly opt into. Everything else in the docs is a walkthrough of how to wire up a specific channel — back to the docs index.