v2.0 — group rooms with opt-in AI

Chat with AI on your own. Bring it into the room.

A private thread with the assistant when you're heads-down. A room where it joins your team only when invited. One conversation model, streaming end to end, moderated at both edges.

No credit card. Your history is yours — training is opt-in.

launch-war-roomAI: mention only
Message the room… use @ai to bring in the assistant
The idea

Most chat apps bolt AI onto messaging. Confluence starts from the conversation itself — a room is just people talking, and the assistant is a participant you choose to bring in, not a tab you have to switch to.

Inside a room

Three people, one conversation, AI on tap

This is a real room's shape — header, roles, an @-mention, and a streamed reply — not a mockup of a different product.

#product-launch

3 members · AI: mention-only

PDA
P
Priya9:41

Pushed the new onboarding copy — can someone sanity check the tone before it ships?

D
Diego9:42

Reads a little stiff in step 3. @ai what's a warmer way to phrase this line?

ConfluenceAI9:42

Try: "You're in — let's get your first room set up." Keeps the momentum without over-explaining.

P
Priya9:43

That's the one. Shipping it.

Group rooms

Everyone in the room shares one thread — not a separate copy each

Owners set the room up, admins manage who's in it, members just talk. Presence and typing indicators make it feel live; reactions and a per-room topic keep it on-track — including for the assistant, which reads that topic before it answers.

MembersRole
PPriya
Owner
DDiego
Admin
AAmara
Member
Posture: mention-only
Off — a human-only room, the assistant never reads it.
Mention-only — it answers when someone writes @ai.
Auto — a full participant, replying without a mention.
AI, invited

It reads the room before it answers — not just the message

When it's brought in, the assistant sees the same thread everyone else does: the topic, the recent back-and-forth, who's asking. Its reply streams into the room as one message that updates live, so nobody is watching a private answer arrive somewhere else.

What you get

Two products that happen to share a spine

A 1:1 assistant and a group chat platform, built on one conversation model — so a private thread and a twelve-person room behave identically where it counts.

1:1 AI chat that streams

A private thread with the assistant. Tokens land as they're generated, stop mid-flight, regenerate a better answer — the supersede chain keeps the history honest.

AI that knows when to stay quiet

Rooms choose their own posture: off, mention-only, or always-on. In mention-only the assistant waits for @ai, so it augments the conversation instead of hijacking it.

Real group rooms

Invite by link, roles for owner/admin/member, presence and typing indicators, reactions, and per-room topics that shape how the assistant answers.

Moderation that fails closed

Every message is classified before it reaches the model and again before the answer is published. If the classifier errors, nothing ships. Verdicts are auditable.

Rate limits you can see

Fixed-window counters per user, per bucket, persisted in Postgres. Thirty messages a minute, ten AI invocations — tunable without a redeploy.

Realtime by default

The assistant's reply is one row that updates as it streams, so every member of a room watches the same answer appear at the same moment.

How it works

Four steps, no setup ceremony

1

Start a thread or a room

A private AI chat takes one click. A room takes a name — and optionally a topic that tells the assistant what this space is for.

2

Bring people in

Generate an invite link with an expiry and a use cap. Redemption runs server-side, so nobody can join a room by guessing a UUID.

3

Set the AI's posture

Off for a human-only room. Mention-only when you want it on tap. Auto when the assistant is a full participant.

4

Ship the conversation

Search every message you have access to, react, edit, regenerate. History is yours; training on it is opt-in and off by default.

Security

The boring parts, done properly

Membership is the only authorization primitive, and it's enforced in the database rather than in application code that someone will eventually forget to call.

The provider key never reaches a browser

All model calls run inside an Edge Function. The client gets a stream, never a credential.

Row Level Security on every table

Membership is the boundary. Audit tables have RLS enabled with zero policies — unreachable from the anon and authenticated roles entirely.

Private attachments, scoped by room

Files live in a private bucket keyed by conversation. Storage policies call the same membership check the tables do.

Opt-in training, off by default

Your conversations aren't training data unless you say so, and you can revoke it in Settings at any time.

Pricing

Priced per person, not per token

Full plan details, limits and the awkward questions live on the pricing page.

Free

For trying the whole thing out.

$0forever

  • Unlimited 1:1 AI chats
  • 3 group rooms
  • 30-day history
  • Community support
Start free
Most popular

Team

For groups that actually ship together.

$12per user / month

  • Unlimited rooms & members
  • Full history + global search
  • File attachments
  • Admin analytics
  • Priority model access
Start 14-day trial

Enterprise

For when procurement gets involved.

Customannual

  • SSO / SAML
  • Audit log export
  • Data residency
  • Custom moderation policy
  • 99.9% SLA
Talk to us

Bring the assistant into the room

Free to start, and the demo runs without an account.

Built by

Chirayu Babu Jaysawal

Crusty-chirayu (Chirayu)