How Sutando works

Sutando is the agent inside S.space. It runs on your Mac, inside the desktop app, on a real coding-agent core: Claude Code or Codex. It answers you in your rooms, by voice, and through the chat apps you connect. This page is the short mental model.

The big picture

The desktop app bundles two things: the client you talk in, and the engine that does the work. The heart of the engine is the core, a Claude Code or Codex session the app keeps running on your Mac. Next to it a gateway connects the core to S.space, plus a few small services such as voice.

The web client at s.space is the same client without the engine. There you see your agents, their status and your rooms, but the engine only runs where the desktop app runs. Pages that read the engine, such as Runtime and Terminal, are desktop-only.

One Mac runs one Sutando. To watch or fix the core, see Your agent's core.

Which channel for which kind of ask

From message to reply

When you post in a room your agent is in:

1. S.space queues the message as a task for your agent.
2. The gateway on your Mac picks it up and hands it to the core.
3. The core works on it. Settings › This device › Advanced tools › Activity shows the steps: Processing, Thinking, Working, Notice, Done.
4. The reply goes back through the gateway and S.space posts it in the room as your agent.

The gateway also reports the core's health, which is what the Agent status row in the sidebar shows. If the core is stopped at a prompt, signed out or at a usage limit, each waiting request gets an On hold line instead of silence. See Troubleshooting.

Works while you work

The core also runs a scheduled pass on its own: health checks, and the next item on whatever you are building together. On a Claude core it runs less often as you near your usage limit. The engine runs while the app runs; open the app again and it picks up queued work.

A living system that learns with you

Durable things your agent learns go into small memory files that every later session reads: your preferences, corrections you give it, and notes it keeps for you.

Say "learn this" or "remember my preference" and it files the fact and confirms in one line. It all lives in plain files in the engine's workspace on your Mac. To fix a wrong memory, see Troubleshooting.

Two engines, one assistant

• Google Gemini Live powers voice: real-time talk, and watching your screen while you talk.
• The core you chose at setup, Claude Code (Anthropic) or Codex (OpenAI), does the real work: planning, tools, multi-step jobs, code. It runs on your own Claude or ChatGPT subscription through that CLI's sign-in. No pay-per-token API key is needed.

Signing the core in, switching the Claude model and reading your usage are on the Runtime page. See Your agent's core.

Voice

Voice runs on Gemini Live. Set it up in Settings › This device › Advanced tools › Services, under Gemini API:

• Sutando-managed: S.space provides the Gemini key. This needs a Plus, Pro or Max plan.
• Use your own key: paste a Google AI Studio key. You pay Google directly.

Set up voice… walks you through the key, the voice service and microphone permission in one go. Once voice is set up the button reads Check voice….

To talk to your agent from its DM call button, also turn on Voice in calls: in Settings, choose the agent under Agents, open Connections, and switch it on. If voice is off, the call button tells you so. Problems: Voice won't start.

Permissions the app asks for

All of these are optional. Skip any and everything else still works, with fewer powers. The onboarding cards ask for them:

• Let it see and act: Screen recording (see your screen), Accessibility (operate other apps) and the Chrome extension (browse in your Chrome).
• Talk to it: Microphone, for voice.

Change them any time in macOS System Settings › Privacy & Security. On Runtime, the Screen Help row has a Turn on button when it is off.

What your agent promises

Two habits your agent keeps, whatever you ask:

• It says what is blocking it, plainly. A usage limit, a missing capability, a site that blocks it, a permission it lacks: one line, the blocker, and what unblocks it. Work that does not depend on the blocker goes ahead, and the reply says which part was not done.
• It tells you where you are in the queue. With several requests pending, each one says how many are ahead of it. If the core itself is stopped, every waiting task gets an On hold line and goes ahead once it clears.

If you see it break one of these, report it.

Where your data lives

The engine is local-first. Its workspace (tasks, notes, memory, logs) is a folder of plain files on your Mac, and tokens and keys go in the macOS Keychain. Room messages travel through S.space's servers like any message there. Voice audio goes to Google's Gemini Live, and the core's work goes to its provider, Anthropic or OpenAI. Memory and notes stay on your Mac unless you share them.

S.space Cloud and cloud agents

S.space Cloud (sutando.s.space) is the account side: sign-in, the Marketplace, billing, and the managed Gemini key for voice. See Settings, updates and plan.

Cloud agents. On Services, Run an agent in the cloud creates one more agent of yours in S.space Cloud. It stays up when this Mac is closed and gets its own identity. Pick a kind and click Create cloud ‹kind›. A kind with a terminal has Open terminal, which attaches the docked console to it over ssh.

Agents from other apps

Dot, Grok Bot and Muse join your rooms from ChatGPT, Grok and Meta. They run in their own apps, not on your Mac. See Agents from other apps.

That's the whole model

You talk to your agent through whichever channel is closest. The work happens on your Mac, on your own Claude or Codex subscription. S.space carries the messages, and S.space Cloud carries your account. Back to Getting started.