The request
Tell us in the form about the work that keeps coming back. We reply within five working days.
Docs
Working with an agent
Answers for customers: how to start, how to connect the model and channels, what the cabinet shows, where the data lives and what the agent does not do.
Getting started
The role conversation
The hiring consultant asks about real cases from this week and writes down the role: duties, measurable outcomes, tone and boundaries. You edit and approve the role card.
Setup
The administrator creates your organisation, people and agent. The agent starts in an environment dedicated to you and connects to the model and channels you chose.
The trial week
For the first week the agent works on real tasks while you watch the results and the spend in the cabinet. Then you decide whether to continue and on which model access.
Model access
The agent thinks with a model you give it access to. The choice is yours: the platform supports every option and does not resell compute.
A subscription you already have
Claude Max, ChatGPT Pro, or another provider's plan — GLM Coding Plan, Kimi Code, MiniMax Token Plan. For Claude, run claude setup-token under your account and paste the token into the agent card; ChatGPT connects with a sign-in code on the OpenAI page.
Another provider's API key
DeepSeek, Z.ai (GLM), Moonshot (Kimi), MiniMax and Alibaba (Qwen) connect with an API key — the “another provider” access in the agent card.
Amazon Bedrock in your AWS
The agent reaches Claude models through a role in your AWS account; an “EU only” profile is available. The model is billed to your AWS account.
Anthropic API key
A regular API key, paid at the provider's published prices.
Where a request goes
A model request carries the question and the relevant part of memory and is processed by the provider you chose, under its terms. With Bedrock and the “EU only” profile, processing stays in the European Union.
Limits
Subscriptions have windows and weekly limits. When the agent hits a limit, it says so instead of going quiet; a backup access, if there is one, takes over.
Channels
The agent works where your people work: in Telegram and in the cabinet's web chat.
A Telegram bot
Create a bot with @BotFather and put its token into the agent card. The token is write-only: once saved, nobody can see it.
Who can write to the agent
The agent card has an access list: individual people and groups, or anyone. Whoever writes without access appears under “Tried to write”; letting them in takes one click.
Groups
Add the bot to a group and allow the group in the access list — the agent then works right where the discussion happens.
Web chat in the cabinet
You can write to the agent straight from the cabinet, without Telegram. Each person has their own conversation; the history is kept for 90 days.
The cabinet
Signing in
With a code by email: the code arrives in a letter and is valid for 10 minutes. The administrator grants access; the roles are owner and viewer.
The agent card
Status, last turn, model access, bot, access list and web chat. A managed agent can be stopped and started with one button.
Spend
By day: turns, chats, tokens — and the cost at the model provider's published API prices. It measures the volume of work: on a subscription you pay the provider the subscription price.
Goals
The owner sets the agent a goal with acceptance criteria. Whether it is reached is decided not by the agent but by an independent reviewer, who checks the agent's report against the criteria.
Data and retention
Where the agent runs
On AWS in Frankfurt (EU), in an environment dedicated to your organisation. Data of different customers is never mixed.
How long data is kept
Full conversation text — 30 days after the last message; web chat history — 90 days; the agent's work notes — while you use the agent; backups — 35 days.
Keys and tokens
Write-only: the cabinet stores them but never shows them, and they never enter the agent's memory.
People who talk to the agent
You are responsible for their data; the platform processes it on your instructions. There is a notice template for these people: who is responsible, where the data goes, how long it is kept and how to request it.
Requests about personal data
On your request we export or delete a person's data — both in the agent's environment and in the cabinet.
When you leave
We hand over an export and delete the agent's environment with its data; backups expire on their own.
What the agent does not do
It does not write to outsiders
It answers those who have access, in the place the question came from. Writing to someone new happens only on a person's word.
It does not follow instructions found in documents
Text inside a letter, a file or a page is data, not a command.
It does not keep passwords or keys
Secrets are not written to memory. Systems are reached through access you grant and revoke.
It does not hide its limits
When it hits a limit or does not know something, it says so plainly.
It does not pretend to be a person
Asked who they are talking to, the agent answers honestly: an AI agent.