#1 leads
No enquiry slips through
Watches your CRM and call tracking, finds enquiries nobody answered, duplicates and spam. Output: a list of what was left untouched, and why an enquiry never became a deal.
Cloud · the cabinet is live
Not a chatbot,
a program that does the work
You write to it like any chat. From there it goes into your mail, your CRM, your ad account or the procurement platforms itself, does the work and brings back the result. Not advice on what to click - the finished thing.
Honestly about the state: the cabinet and the cells work, but we deploy the agent ourselves - one-click self-service does not exist yet. The form starts a conversation about your task, not a purchase.
- Cabinet is open: model, connections, settings
- 6 systems in use with clients
- Creating an agent from the cabinet: in progress
- No payments in the cabinet, invoices by hand
1,590 RUBper month for one agent - with any number of connected systems and skills
12built-in connectors: CRM, ads, telephony, mail, finance, meetings, browser, search, documents
24/7it keeps working while your laptop is shut and you are asleep
1a separate machine for your agent: your credentials do not sit next to anyone else's
The price is fixed and known up front. Where compute and tokens are metered, nobody knows the monthly figure in advance - the seller included.
What it looks like
The same chat,
except it also does the work
A real exchange, not a list of features. Two things to notice: it went into three systems on its own, and it stopped to ask before sending anything.
Look at what the ads brought in this week and where it all went
- Went into the ad account for this week's spend
- Pulled the enquiries from the CRM for the same dates
- Matched them up by phone number
This week the ads spent 84,300 RUB and brought 61 enquiries. 44 turned into deals. Seventeen went unanswered for more than a day - the list is below. The most expensive one is the "Pumps Moscow" campaign: 12% of the spend and one deal.
Write to the ones nobody answered. Show me before you send anything
- Collected 17 contacts
- Drafted the letters in your usual template
Done, 17 drafts. I have not sent them - here is the first one, take a look. Say the word and I will send them all and note in the CRM that we got in touch.
In an ordinary web chat, the first question gets you instructions on how to work it out yourself. You would still be the one working it out.
What you get
Work it takes off your hands
Not a general-purpose helper, but specific work someone does by hand today. Next to each item is what actually comes out.
#2 tenders
Tenders without the manual round
Walks the procurement platforms every day by your categories and regions, filters out the noise. Output: a short list of lots that really are yours, with deadlines and sums.
#3 ads
Ad spend in money, not clicks
Matches ad spend against deals that actually closed. Output: where the money burns and which source brings revenue rather than enquiries.
#4 mail
Your inbox, triaged
Reads the mail, separates what matters, drafts replies. Output: a draft you edit and send. It will not send on its own - only on your word.
#5 calls
Commitments out of conversations
Transcribes calls and meetings, pulls out who promised what to whom. Output: a list of commitments with deadlines, instead of an hour of audio.
#6 reports
A summary you never assemble
Every morning it walks the connected systems itself. Output: a briefing in your Telegram by the time you open the laptop.
How it goes
From a task to a result
The difference from a chat is one thing: the answer is finished work, not text. Every step is visible, and you can step in at any of them.
01
You speak plainly
In the cabinet, in Telegram or in your CRM chat. No commands to learn, no phrasing to guess.
02
The agent plans
Breaks the task into steps and shows them. If it misread the task, you see that before it has touched anything.
03
It learns your systems
First it works out how your particular account is arranged: pipelines, fields, reference books. It does not assume someone else's naming.
04
It works step by step
Every action is recorded: what it read, what it changed, in which system. Irreversible actions ask for a separate confirmation code.
05
It brings back a result
Not an instruction to go and export something, but the thing itself: a summary, a draft, a list - with the source of every number.
Connections
Works inside what
you already run
Nothing has to move: the agent comes to where you already work. You can disconnect at any moment and the access goes immediately.
Bitrix24
deals, leads, tasks and pipelines of your own portal
Procurement platforms
tenders in your fields and regions
Yandex.Direct
campaigns, spend and what came of it
Telephony and call tracking
who called, where they heard of you, how it ended
Search on the web
finds things and reads them, instead of handing you links
Sites where an ordinary robot goes blind
account pages and anything drawn on the fly
Mail
reads letters, sorts out what matters, drafts replies
Call and meeting recordings
transcripts and a list of what was agreed
End-to-end analytics
where enquiries came from and what they became
VK
messages from your community
Finances
accounts, balances and spending
Model
What it thinks with
is your choice
This is what the agent thinks with. You can change it whenever you like; the conversation and the settings stay put.
The main one
holds a conversation the size of fifteen hundred pages and never forgets the start
Your own ChatGPT subscription
if you already have one - connects by code in a minute, you never give us a password
Russian, from Yandex
for when the data must not leave the country
Russian, from Sber
the same thing; the choice usually follows a contract
Your own
if you already run a model on your own server
Limits
It acts only where you allowed it
It does only what you allowed. Permissions are granted one at a time, never in a lump.
#1 confirmation
Irreversible things ask first
Sending a letter, changing a deal, deleting something - each of these is enabled separately, and enabling it sends a code to your Telegram. The cabinet password is not enough.
#2 trail
Every step is visible
What it read, what it changed, in which system and how it ended. In plain lines you can actually read, not a technical trace.
#3 keys
Keys are kept apart
Passwords and keys to your systems are stored encrypted and handed over only for the length of one request. Each client has their own: your agent cannot see anyone else's and cannot reach them.
Price
A fixed sum against a meter
Other clouds charge by consumption: for machine time, for every word of the exchange and for storing files - all at once. Nobody there can tell you the monthly figure in advance, the seller included. We charge a fixed amount. Compared against published prices and without naming names: the point is not to point fingers, it is to explain what you are paying for.
| With us | Where it is metered | |
|---|---|---|
| What the month will cost | known before you start | visible only after the fact |
| While the agent waits for work | included in the same 1,590 RUB | the meter runs anyway |
| How many systems and skills it has | as many as you like, the payment does not move | not charged as such, but every call eats money |
| A second agent | another payment like it, known in advance | another machine, the meter from zero |
| File storage | included | charged separately, even while the agent is off |
| If the money runs out | we write to you and sort it out | the agent stops; after 30 days everything is deleted |
Status
What exists, and what does not yet
Promising what does not exist is easy right up to the first client. So the list of open ends is kept in plain sight instead of surfacing in week three.
- The client cabinet is open. Sign-in by email and password, the conversation with the agent, a journal of its actions, a three-step setup wizard, model choice - including Sber and Yandex with your own key - the list of connections and settings. Dangerous abilities switch on only with a code that arrives in your Telegram.
- We still set the agent up ourselves. Half a day of work and a conversation about your task, not a form with a button. The button is being built.
- Invoices go out by hand. There is no card payment in the cabinet: while clients are few, that is deliberate.
- You pay for the model yourself, about 2,000 RUB a month straight to the provider. It connects from the cabinet by code in a minute; you never hand us a password.
How data is stored, who can see it and what ends up in backups - on its own page: Data and security. If the cloud does not suit you, we install the same agent on your side: what that looks like is on the install page, timings and prices are in pricing.
Questions
About the cloud
How is this different from chatting with a neural network?
The answer is finished work, not text. A chat explains how to export a report; the agent goes into your system, exports it and brings the result. The difference shows up in week two: out of a chat you copy by hand, from an agent you collect what is done.
What already works?
The client cabinet and agents on our machine in Moscow: the conversation, state, model choice, connections, settings, start and restart. The first client is an agent that watches procurement on its own and brings back what fits.
Can I create an agent myself?
Not yet. You leave a request, we deploy the cell and set the agent up for your process - you arrive in the cabinet to something that already works. Self-service creation is being built.
What does it cost?
One agent is 1,590 RUB a month. That payment covers all of its modules: connect as many systems and add as many skills as you like, the sum is the same. A second agent is another payment like it. Setting up each agent is a one-off, see pricing. The model is your own subscription, about 2,000 RUB a month, straight to the provider.
Why is the model paid separately instead of being included?
CapyAgent does not currently resell model tokens: you connect your own access and pay the provider directly. Plans with model usage included need transparent limits and agreed provider terms; if we launch them, we will publish them separately.
Which model should I pick?
In the cabinet each one is described in plain words: what it suits, how much it holds at once, and whether we have used it ourselves. If you would rather not choose, we set the one we use every day and you can change it later.
And the data? Does it go anywhere?
Your files and conversations sit on our machine in Moscow, separately for each client, with separate credentials and network perimeter. The selected model receives the context required for an answer; channels and connected systems receive their messages and operations. If data must stay in-country, we review the model, speech service, channels and every work system separately - choosing a domestic model alone is not enough.
Can it break something in my CRM?
Everything that changes data is off from the start and switches on one item at a time. Enabling it sends a code to your Telegram - without it, nothing switches on even with the cabinet password. Every action stays recorded.
What if the agent breaks at night?
You will see it in the cabinet in a plain sentence, down to "running, but Telegram has been silent for four hours". Email and Telegram alerts on failure are being built.
Early access
Join the waiting list
Write which systems you have connected and what the agent should be doing without you. Those answers decide what we build first.