It doesn't suggest. It acts and closes the loop.
Every request becomes a plan. The orchestrator picks ONE system per round and hands a self-contained task to a specialist agent, which runs the real MCP tools and reports what it did. The cycle has a budget — four delegations per message — and always lands on an answer.
Every round ends: an empty reply becomes another attempt, and the round limit answers with whatever was found.
The trajectory stays in plain sight — open "View trajectory" and read every tool that was called.
An unreachable connection contributes zero tools: the chat degrades to what is left instead of breaking.
The conversation
One channel, and you never have to organise it
No creating conversations, no picking threads, no losing your place when the tab closes.
Continuous chat
You just write. The assistant classifies on its own whether this continues the thread or opens a new topic, and archives the previous one into History when the subject turns.
History and resume
Closed conversations stay read-only, and scrolling up loads the earlier ones. Closed the tab mid-answer? It picks up where it left off when you come back.
Live answer, no stopwatch
The text appears as the model writes it. The round runs outside the API's 29 s ceiling, so even a slow local model on your own machine gets to finish.
Per-message trajectory
Every answer keeps its own machine steps, folded away. One click opens the list: which agent was called, with which tool, and what came back.
Acting on your products
It operates your products — with your access
No "ask your administrator". The tools run with your token, and the target product checks.
One agent per system
Each connected system has a specialist agent that only sees its own tools. The orchestrator delegates the task and demands the result — small decision, small mistake.
Catalog read live
Tools are read from the product on every request, never from a frozen list. A new S7 Calendar tool shows up in the very next conversation.
Automations on demand
Save a recurring prompt with a name and a schedule and run it whenever you want, with the same tools as the chat. Every run is recorded: outcome, duration and who fired it.
Always with your access
Every tool runs with your token and the product re-checks your role in the Space. The assistant never does what you couldn't do yourself.
Systems in this phase's catalog
Memory
It remembers what matters and writes down what happens
Two different memories: what it learned about you, and the diary of what happened in the Space.
Memories
When a topic is archived, the assistant distils facts, preferences and decisions into short notes. The most recent ones join the next conversation — and you can edit or delete any of them.
Notebook
A markdown document organised by date that the assistant writes and reorganises. Once a day it re-reads the notebook, compacts the old and keeps what still serves.
Listening to products
Turn on S7 Tasks and S7 Calendar and their events become notebook lines under today's date. Deterministic logging: no AI, no cost.
Voice, images and presence
Talk in whichever way is fastest
Text, audio or an image — and an assistant that looks like your Space.
Voice message
Record instead of typing. The audio is transcribed before the round and the text joins the conversation as if you had written it.
Spoken replies
Play any answer with one tap, or turn voice mode on and hear them all. The audio uses the voice described in the Space's persona.
Images and attachments
Paste, drop or attach images and audio — up to four per message. The image is described before the round, so the assistant reasons about what is in it.
Persona and presence
Give your Space's assistant a name, a tone, modes and colours. Each machine picks the presence — Constellation, Aurora or Nebula — and Ambient mode makes it react to your music through the microphone, recording nothing.
How it works
One round, from the ask to the done
The request becomes a job stamped with your token, and that job only disappears when it finishes — with the result or with the reason. Nothing runs on anyone else's behalf and nothing hangs in silence.
You ask
In text, voice or an image. The assistant decides whether this is the same topic or a new one before anything else.
It picks the system
The orchestrator reads the live catalog of connections and hands a self-contained task to a specialist agent.
The agent executes
The agent runs the product's MCP tools with your token; the product re-authorises your Space access before accepting.
You get the result
The answer says what was done, and every step of the way stays in the transcript — including the attempts that failed.
What holds it up
Built for a Space, not for a text box
The same guarantees of the S7 platform apply inside the assistant.
Space and roles
Everything lives inside a Space, and every route re-checks your membership on the server. A viewer can read the assistant, but cannot act through it.
The conversation belongs to the Space
Every conversation belongs to the Space: any member reads it, viewers included. To answer, the content goes to the AI provider configured in the Space's Relay — and it does not become an event, does not reach logs and never crosses the bus.
Inference through S7 AI Relay
The assistant asks for an activity and an effort, never for a model. The Space's Relay routing decides — a cloud provider or a local model over the bridge — and every token is metered in your Space.
Three languages, end to end
Interface, error messages and even the assistant's internal prompts in pt-BR, en-US and es-ES.
Under construction
S7 AI Assistant is still being built. What does not exist yet is listed here, with no promised date.
Automations on a schedule
The schedule is already stored, but autonomous firing waits on the platform automation credential. For now, runs are on demand.
More connected systems
In this phase the connection catalog is fixed: S7 Calendar, S7 Tasks, SNIC, S7 Trail and the assistant's own automations.