An AI that knows your business and asks before it acts.
Kubrik Copilot lives inside Kubrik OS and Kubrik POS. It answers with your real data, explains where the numbers come from and prepares changes for a person to confirm. It uses your own permissions, never more, and it never touches the books on its own.
Example conversations using the real tool names. The data is illustrative.
Where you already work, and from your own AI too.
A tabbed panel, one module per tab
Open several conversations at once, each focused on a module: catalog, stock, orders, accounting, customers, HR, projects or documents. A tab can coordinate several modules with playbooks.
Ctrl+K at the till and in the back office
The cashier asks for cart changes in plain words and applies them with Enter. The manager diagnoses stuck vouchers and payments. It never charges or invoices on its own.
MCP with personal keys
Connect any MCP-compatible client to mcp.kubrik.io with a personal kos_ key. It works with your permissions and writes come back to Kubrik as proposals to confirm.
The rules live in the code, not in the prompt.
A language model can be wrong or be tricked. So Kubrik doesn't ask it to behave: it only gives it what the user may do, and everything that changes the business goes through a person.
It can't see what you can't use
Before each turn, tools are filtered by the person's permissions. They're checked again when they run.
It proposes, it doesn't execute
Each change arrives as a card with its details fixed. If nobody confirms it within 24 hours, it expires. Pending ones wait in an inbox.
Mandates with amount bands
You set what it may do alone, what it must ask about and what it may never do, by amount, branch, warehouse or time window. The system computes the amount, not the model.
Accounting, never on its own
Posting journals, paying suppliers, recording receipts or voiding invoices always asks for confirmation, whatever the level or mandate.
Everyone in their own lane
Each company's data is isolated. At the POS, the cashier stays bound to their sale and branch: if the model asks for another, the system refuses.
Everything is recorded
Every agent write, even failed ones, lands in the platform audit log with the agent as author. So does every tool call.
The right model for each question, with a cap you set.
A simple lookup goes to the fast model; a cross-module analysis, to the most capable one. Kubrik records the cost of every message, warns at 80% of the monthly cap and, if you want, stops there. You can use your own Anthropic key.
It remembers what matters, and you can review it
Memory stores typed statements with a source and an expiry, filtered so personal data isn't kept. You can see and delete it from Kubrik OS.
It talks the way you talk
It answers in the asker's language and adapts to the company's country: voseo in Argentina, tú in Spain and Mexico, US English and Brazilian Portuguese.
Ask Kubrik from the AI you already use.
Kubrik exposes 295 tools from 11 modules over the open MCP protocol. Each module adds a tool to talk to its agent. Your AI client works with your permissions, and what it writes comes back to Kubrik as a proposal you confirm.
Work that runs on its own, on a schedule and a budget.
Schedules and triggers
An agent runs at the local time you choose, or when a financial signal calls for it, always with a budget per run.
Briefings with citations
Every statement in a briefing says whether it's a fact, an observation, a hypothesis or a recommendation, and its citations are checked against the data.
Plans to approve in one go
A run's proposals are grouped into a plan you approve at once. An agent can delegate to another, with depth and budget limits.
Add an AI that works by your rules.
Kubrik Copilot comes with Kubrik OS and Kubrik POS. See the detail on levels, tools and security, or try it on Kubrik Cloud.