For whoever decides and whoever implements

Kubrik Copilot, in detail.

Autonomy levels, mandates, tools by module, background agents, models and costs, memory, external MCP and the rules the system always enforces.

Autonomy levels

Three levels, per person and per module.

The administrator sets each person's level in each module. Every tool also carries a risk tag: read, safe, reversible, destructive or playbook.

Consultant

It only reads and explains. Tools that write aren't even offered to the model.

Executor

It can propose changes, and each one is shown as a card the person confirms before anything happens.

Autonomous

Only safe actions run on their own: notes, drafts and memory. Reversible and destructive ones still ask for confirmation.

Mandates

Written rules for each agent.

A mandate says, for each action, whether the agent may do it alone, must ask, or is forbidden. A mandate can only narrow what the level and the tool's tag allow; never widen it.

OutcomeWhat happens
AUTORuns without asking, within the limits
CONFIRMBecomes a proposal for someone to approve
DENYNot allowed, not even with confirmation
ConditionsSystem-computed amount, branch, warehouse or company, time window and a daily cap on automatic actions
ApprovalBy role, and wired into the Kubrik OS approval engine
Tools

347 tools, counted one by one.

Each tool is named module_action. 295 can be used over MCP from outside; platform, playbook, memory and POS tools only work inside Kubrik.

ModuleToolsOver MCP
HR45yes
Professional services42yes
Warehouse40yes
Product catalog37yes
Accounting36yes
Orders34yes
CRM32yes
Documents20yes
ModuleToolsOver MCP
Customer analytics3yes
Purchasing planner3yes
Finance (dashboard)3yes
Kubrik POS14POS only
Platform25no · all need confirmation
Cross-module playbooks7no
Memory, plans and delegation6no
Total347295

wms_trace_stock_out_root_cause

Explains why a product ran out, with the movements that took it to zero.

pms_trace_sku_readiness

Says what a SKU is missing to be ready to sell on a channel.

acc_trace_invoice_aging

Explains the aging of an invoice that is still open.

rpos_trace_fiscal_document

Diagnoses a failed ARCA voucher and, with confirmation, proposes retrying it.

pur_planner_propose_reorders

Builds purchase proposals with the Purchasing planner.

psa_get_leakage_report

Shows the hours that were worked and won't be billed.

Cross-module playbooks

Several steps, several modules, one confirmation.

A playbook builds a full operation across modules and presents it whole to confirm. Playbooks always ask for confirmation.

orch_lead_to_quoteFrom a lead to a quote
orch_onboard_customerFull customer onboarding
orch_provision_productProduct family and its first SKU
orch_replenish_low_stockFrom low stock to a purchase proposal
orch_create_and_dispatch_orderCreate an order and dispatch it
orch_confirm_and_ship_orderConfirm an order and ship it
orch_bulk_applyApply the same change to many records
Background agents

Agents with a schedule, a budget and accountability.

Runs and schedules

Each agent runs at the cadence and local time you set, or when a financial signal triggers it. Each run has its own budget.

Verifiable briefings

Each statement declares whether it's a fact, an observation, a hypothesis or a recommendation, and its citations are validated.

Delegation with limits

An agent can ask another, up to two levels deep and with up to 60% of its remaining budget. Delegated work can only propose.

Plans

A run's proposals are gathered into a plan approved in a single act.

Financial advisor

An agent that builds a daily briefing on the business's financial health and the signals that changed.

Agent catalog · on the way

Ready-to-hire agents, like a collections analyst or an inventory watcher. The catalog exists, but hiring them isn't enabled yet.

Models, budget and memory

What happens under each answer.

Model router
Simpleclaude-haiku-5-5
Mediumclaude-sonnet-5-5
Complexclaude-opus-5-5
How it picksIt scores each turn by level, modules involved, tools, intent and conversation length. Within a session it never steps down.
Per companyAutomatic or pinned, with a tier ceiling and your own Anthropic key if you want
Budget and quotas
Monthly capIn cost and in number of requests; warning at 80%; block or only warn
RecordThe cost of every message is stored
Per-turn limitsUp to 50 tools and 6 rounds per turn
SessionsUp to 5 open conversations per person
Memory
What it keepsGoals, policies, preferences, decisions, outcomes, seasonality, known risks and notes
ScopeCompany, agent, or agent and person
CareSource, status and expiry; personal-data filter; review from Kubrik OS
Language
AnswersIn the asker's language, with output verification
RegisterVoseo in Argentina; tú in Spain and Mexico; US English; Brazilian Portuguese
External MCP

Your AI client, with your permissions and no more.

The effective level is the lowest of the key's, the person's and the company policy. Below Autonomous, every write comes back as a proposal with a link to confirm it in Kubrik OS, and refusals are audited too.

Addresshttps://mcp.kubrik.io/<module>
ProtocolMCP · Streamable HTTP · JSON-RPC (initialize, tools/list, tools/call)
KeysPersonal kos_ keys, up to 10 active per person, with expiry, renewal and revocation
ModulesCORE · OMS · WMS · CRM · CRA · PUR · HR · ACC · DMS · PSA · DASH
Agentask_agent in each module runs a turn of its agent
Security and hard limits

What the system enforces, no matter what.

Security

  • Tools are filtered by permissions before they're offered to the model and checked again when they run.
  • Data isolated per company; at the POS every tool checks the branch and the cashier stays bound to their sale and till.
  • Every agent write lands in the platform audit log with the agent as author, including failed ones.
  • A master switch turns the copilot off instantly if needed.

Never without a person

  • Posting journals, paying suppliers, recording receipts or voiding invoices: always with confirmation.
  • At the till it doesn't set prices, discounts or approvers, doesn't charge and doesn't confirm the sale. If the cart or a price changed, the proposal is refused.
  • It doesn't issue fiscal vouchers: it only proposes retrying an existing one, with confirmation.
  • Platform tools are all destructive and never run on their own.

On the way

These aren't available yet. We list them so you know what's coming.

Hiring agents from the catalog Pricing, promotions and loyalty tools Copilot in the mobile apps Assisted bank reconciliation Document reading

Try it with your data.

Kubrik Copilot comes with Kubrik OS and Kubrik POS. Start on Kubrik Cloud or go back to the overview.