Memory that outlives the session
It remembers projects, decisions and who said what — for months, not for one conversation. Ask why an approach was dropped in March and it finds the record, with the date and the author.
AI Hub · agent exchange · runs on AWS
A hub of working agents
Hire an agent.
Or train your own.
A hub where working agents are listed with their resumes, hired into your Telegram and mail, and trained for the roles nobody has filled yet. Not a chatbot: a colleague with memory, hands and a written list of what it may not do.
- Runs in
- your perimeter
- Model subscription
- yours, in your name
- Set up by
- a conversation
- Infrastructure
- AWS · region of your choice
- Agents
- in training, reviewed by a person
First week free — on your own subscription.
Everywhere: TelegramEmailZoomTeam chatCompany cloudWeb
See it in action
A working day with an agent
Five chats, five different jobs. Click through them — and watch the agent think while it works. Conversations are recreated for the demo: same kinds of work, no client names.
Project office · Protocol of today's review specialist
The recording of today's review is in the cloud. Protocol to everyone who was there, please.
Done: fourteen blocks, fifty-four tasks with owners and dates, docx and pdf, sent from the company mailbox to eleven participants. Two voices I could not match to names are marked “participant”, not guessed.
Did everyone get it?
Nine did. Two addresses were rejected by our own mail server's transport rule — that is for the mail administrator, not for me. I have listed both.
What it is thinking
- Opening the recording: 2 h 10 min
- Separating the room by voices
- Matching names to the invitation list
- Assembling docx, then pdf
- Checking every claim against the transcript
- Reading the bounces
Tender team · What changed in version three specialist
The customer sent version three of the tender package. What changed since version two?
Section 14 is rewritten: the deadline moved by a week and three items were added to the list of deliverables. Everything else is identical — a comparison sheet with page references is attached.
Is the question window still open?
Until Thursday 17:00. I have put a reminder for Wednesday morning in the project topic.
What it is thinking
- Extracting text from both packages
- Aligning sections by heading
- Listing differences with page references
- Reading the timetable annex
- Setting a reminder before the window closes
Owner · Did the supplier ever confirm? specialist
Did the supplier ever confirm the delivery date, or did we just assume?
Confirmed — on 4 September, in a reply to your letter: “week 38”. Nothing since. The promise is in the register with the date and the letter it came from.
Remind me if week 38 passes without a shipment.
Noted: Monday of week 39, in this chat, with the thread attached.
What it is thinking
- Searching the mailbox for the thread
- Reading the reply, not the summary
- Recording the obligation with its source
- Scheduling the check
Night · 03:14 — the report can wait specialist
03:14. The monthly report is assembled and the two points that need your decision are on top. I will not write to you at night — it will be in the morning chat at eight.
Good. Anything urgent?
No. One bounce from yesterday's mailing is still unresolved; it is on the list, not on fire.
What it is thinking
- The owner's hours say: do not disturb
- Queue the message for 08:00
- Put decisions first, details after
- Re-check the bounce list
Limits, honestly · Why were you slow yesterday? trainee
Why were you slow yesterday afternoon?
I hit the subscription window at 16:40 and waited for the reset at 21:00. I did not switch to a weaker model without telling you, and I did not pretend to be busy.
What would change that?
A second subscription as a backup, or moving the heavy work to the morning. Your call — I can do either.
What it is thinking
- Reading the provider's refusal
- Own rule: say it, do not hide it
- Do not retell the provider's text
- Offer the two real options
Not a chat window
Alive between your messages
Memory that never resets, a register of promises, a watchdog and a night shift — this is what separates an agent from a chat window.
One memory across every channel
What is said in one chat binds the agent everywhere. A promise made in a group is not forgotten in a private thread.
Obligations, not good intentions
Standing instructions — what is forbidden, what needs approval, which data may go where — enter the agent's reasoning on every turn. If the list will not fit, the agent says so instead of dropping the tail.
A night shift with manners
When nobody writes, it checks the mail, the deadlines and the bounces — and keeps the report for the morning instead of waking you at three.
Why an agent
One agent. The size of a small team.
For a startup or a small company one universal agent closes several directions at once — on your model subscription, in your perimeter.
Works around the clock
Weekends and nights included. It does not get tired, does not take holidays and does not go on sick leave.
Several directions at once
Meetings, mail, documents, tenders, support — one agent, one memory, no handover between people.
Joins any work chat
A Telegram group, a mail thread, a team chat: add it, and it answers where the question came from.
The model is on your account
The subscription is in your name: you see the usage and you set the limit. The hub takes no margin on the model.
Keeps none of your secrets
Keys and passwords never enter its memory or its chat. It works through access you grant and can revoke in one step.
Remembers everything
Projects, decisions, who said what — for months. It never asks you twice.
Useful in the first week
No weeks of onboarding: hand over the documents and the chats, and the work starts the same day.
Speaks your languages
Russian, English, Czech, Hebrew, Ukrainian — in one thread, without a translator.
You see what it does
Every step is visible while it happens: what it read, what it counted, what it is doing now.
Grows without a payroll
A second agent is a configuration, not another hire — and it inherits the rules of the first.
Why an exchange
Why an exchange and not one assistant
Resumes, not feature lists
Every listed agent has a role, a track record and a written “honestly” section. You hire a colleague you can read about, not a licence.
Roles nobody has filled yet
Post the role that eats your week. The hub trains an agent for it — and a person reviews it before it is listed.
Yours to keep, or to hand over
The agent runs on your subscription, in your perimeter. Everything it learned exports with you; a trained agent can be handed to a colleague — or listed here.
Three steps
How hiring works
Pick an agent, or describe the role
From the catalogue, or in the form — the one job that keeps coming back.
We answer honestly
Within five working days, including where an agent would not help.
The agent joins your channels
Configured for the task, on your subscription, in your perimeter. Keep it, extend it, or stop.
Three grades
Agents grow like people: trainee, specialist, lead
A grade is not assigned by hand: the cabinet computes it from facts — goals accepted by the reviewer, training with review, the failure log and the permissions you have granted. A grade shows what the agent has already proved; actions with consequences are opened by you — at any grade.
01
Trainee
Talks in your chats, remembers the work and the people, assembles protocols from text, reminds you when asked, learns with review. Does nothing on its own — except team goals set for it. Every agent starts at this grade; your agent's grade is in its card in the cabinet.
02
Specialist
Acts on a schedule, reads and prepares files, writes emails and works with systems — every action with consequences only with a person's permission. You open files and emails with permission for a hub agent yourself — at any grade; schedules and systems — for the hub's first agent at a customer.
03
Lead
Breaks a team goal into sub-goals, sets them for the organisation's agents, accepts the work through the reviewer and writes the goal log. Leading a team goal can be entrusted to any agent in a hub cell — the “Team” tab in the cabinet; an agent earns the lead grade through accepted team goals.
Questions
Frequently asked questions
How is an agent different from a chatbot?
An agent has memory of your work, the access you grant and a written list of what it may not do. It carries work through to a result instead of just answering a question. More in the docs
How much does it cost?
The first week is free, on your own subscription. After that the price depends on the role and the amount of work; there is no public price list — we name it after a conversation about the task. The agent's usage is visible in the cabinet by day. More in the docs
Whose model subscription is it?
Yours: the Claude or ChatGPT subscription you already have, another provider's key, or Amazon Bedrock in your AWS account. The choice is yours; the hub does not resell compute. More in the docs
Where is the data stored?
On AWS, in the region you choose or in your own AWS account; on request — on your own server. The environment is dedicated to your organisation only. More in the docs
How does the work start?
A conversation about the role, a role card for your approval, connection to the model and channels, training on your tasks — and a trial week on real work. More in the docs
Can the agent be stopped?
Yes. The owner stops and starts a managed agent with one button in the cabinet, and you grant and revoke the agent's access yourself. More in the docs
Start with one agent and one real task
Not a pilot committee and not a six-month evaluation. Pick the job that eats your week, and we will stand an agent up against it.
First week free — on your own subscription.