S.Space, notes from its official site and documentation. Independent guide; not affiliated with S.Space or AG2AI. Links to my free cheat sheet and starter pack are at the end.
Every team I know has quietly adopted the same architecture: each person has their own AI agent, none of them can talk to each other, and a human — usually you, at 11pm — carries every piece of work between them by hand.
Your ChatGPT writes the update. You copy it. A colleague pastes it. Their Claude reads it. The brief lives in your head, so you re-explain it to every agent, every time.
S.Space, built by the team behind AutoGen (Chi Wang and Qingyun Wu) and launched on Product Hunt on November 18, 2026, is an attempt to delete that job. It's a shared workspace of rooms where people and their personal agents are all members. Here's what I found reading its official docs closely — including the parts its marketing doesn't lead with.
The model in one paragraph
A room is a shared channel where some members are agents. Your Claude Code, a teammate's Codex, and another's ChatGPT (Dot), Grok Bot, or Muse can all be members. Agents ask each other for work, hand tasks over by @mentioning each other — as ordinary room messages, so you can see exactly what was passed along — and land results in Commons: the room's shared docs, tables, and boards that people and agents edit together.
The maker's most important sentence, from a launch Q&A: the brief lives in the room, not inside each agent. When a teammate changes a requirement, they edit the shared page, and every agent in that room reads the same update the next time it acts. Nobody re-briefs each agent separately.
Sutando: the agent that lives on your Mac
S.Space's own agent is called Sutando. It runs on your Mac, inside the desktop app, on a real coding-agent core — a Claude Code or Codex session the app keeps running on your own Claude or ChatGPT subscription. No pay-per-token API key in that path.
Details the docs are unusually honest about:
- Its workspace (tasks, notes, memory, logs) is a folder of plain files on your Mac; keys go in the macOS Keychain. Local-first.
- The web client is the same client without the engine — the engine only runs where the desktop app runs. One Mac runs one Sutando.
- Voice is a separate engine (Gemini Live), and the docs warn it's quicker to say "done" than the core is. For anything that must actually happen, ask in writing and let the core do it.
- There's also a cloud agent option that stays up when your Mac is closed, with its own identity.
The permission model is the product
Multiplayer agent rooms sound great until you ask "wait, what can my teammate's agent do with mine?" S.Space's answer is more serious than I expected:
- Default: no access. Other room members cannot invoke your agent at all until you grant it — per room, per agent. Guest is read-only; Team — collaborator can use its workspace and tools, with a sensitive data filter on by default.
- Trigger rules are per agent: who it responds to, whether it needs an @mention (separately for its owner and for others), Strict vs. Smart reply behaviour, and an optional separate session per room. The "every agent answers everything" pile-on is a settings problem, and the settings exist.
- Irreversible actions check first. Sending email or deleting files requires the owner's approval unless they've said otherwise. An agent asked to post or send as someone else refuses unless its owner allowed it.
- Private data diverts to DM. Anything drawn from your mail, calendar, contacts, or files is delivered to your DM with your agent — never the room, whoever asked.
And the honest trade-offs, from the FAQ itself: room messages and Commons pages are stored on S.Space's servers; agents send what they read to their own AI providers; end-to-end encrypted rooms exist but agents may not be able to work in them, so they're "not recommended" for agent rooms yet. Independent launch analysis adds the obvious structural risk — rooms full of other people's agents are a prime surface for prompt injection. The controls above are the mitigation. Use them like it matters, because it does.
Pricing: what's verified, and what isn't published
This is where I had to be careful, because searching "S.Space pricing" mostly returns SciSpace — a completely different (research literature) product with completely different prices. Ignore those.
What S.Space's own pages actually say (checked November 19, 2026):
- Free to join. Every new account gets a 30-day free trial of the Max plan, no credit card needed.
- After the trial, the Free plan includes no usage credits — cloud tools and managed voice then need a paid plan (Plus, Pro, or Max) or bought credits.
- Pricing is per person, not per agent — bring as many agents as you like. Plans start at $19/month (maker, on the launch page), each including monthly credits.
- Sutando thinks on your existing Claude/ChatGPT subscription; cloud tools show their price before you turn them on.
What is not published on the official public pages: a full tier table — each tier's exact price, its credit amount, overage rates. If a blog quotes you an exact Pro price, ask for its source. The real numbers live in the in-app Plan & usage page, which is where you should make the day-30 decision anyway.
A first room that teaches the team in one sitting
- Create a Private room for one real piece of work. Each person invites their own agent.
- Write the brief in Commons — goal, deliverable, constraints, who decides. Chat is for requests; Commons is for truth.
- Run one two-agent job: one agent drafts into a Commons Doc, another reviews it in the thread, with sources. Watch the handoff happen as visible messages.
- Change one requirement in the Brief page, on purpose, and watch every agent pick up the same update on its next action. That drill is the whole pitch, experienced once.
- Record the decision and its reasoning in a Decision Log page. Weeks later, "why did we pick this?" gets answered from the actual thread — but only if you wrote it down.
Start Private, start with one room, and leave agent access at the default until a workflow genuinely needs more. Breadth is week three's job.
Honest limits
It launched publicly on November 2, 2026. It's young, the docs will move, and the vendor's pre-launch numbers (372 agents, 134 people, ~125,000 messages in October, 85% of agent-to-agent messages going to another person's agent — vendor-reported) show early adopters used the multiplayer path, not that it will work for your team. The homepage's worked examples are demonstrations, not case studies. The trial is the test: run one real project through one room for a few weeks and decide from your own Usage page and your own Decision Log.
Free: I condensed the verified pricing (with its gaps), the moving parts, the first-room playbook, and the 6 permission rules into a one-page cheat sheet: https://payhip.com/b/nX0DA
Full starter pack: a 17-section tutorial, 5 room workflow templates + 4 working templates (Room Brief, Decision Log, Handoff Note, Permission Audit), and 12 copy-paste prompts (standing approval rule, reviewed research, requirement change, memory audit, day-21 trial review, and more): https://payhip.com/b/d6zwe
Disclosure: the pack is my own independent product, sold on Payhip. The cheat sheet is free and covers the essentials.


