An AI agent
inside every
Kubrik product.
Kubrik Copilot lives inside Kubrik ERP, Kubrik POS, and Kubrik Commerce. It diagnoses problems with evidence, edits with preview-and-confirm, and reasons across modules in one turn — powered by 400+ tools and Claude, over the open MCP protocol.
Read. Confirm. Or auto-run safe writes.
Each user gets a tier per module. Tool catalogs filter what the LLM can see before the model decides what to do — the guarantee lives in code, not in prompts.
Read-only diagnostics
The agent answers questions with evidence — what changed, what was rejected, what's missing — but cannot mutate state.
- Trace primitives (readiness, attributes, prices, allocations)
- Cites recent mutations from
ChangeHistoryEntity - Cannot call write tools — they're never advertised to the LLM
Preview-and-confirm writes
Writes become proposal cards. Nothing mutates until the user clicks Confirm. Proposals persist if the tab closes.
- Exact args + dry-run preview rendered inline
- Tools tagged
SAFE/REVERSIBLE/DESTRUCTIVE - Idempotent confirm — double-click = no double-write
Safe writes auto-run
Only SAFE-tagged tools run without confirmation. Reversible and destructive ones always surface a confirm card.
- Notes, drafts, tags can land in one turn
- Auto-action banner in the UI for every silent write
- Audited row in change history with originating message
Three core moves — the rest is composition.
The same engine answers a why-won't-this-publish question, edits a SKU on confirm, and chases an order discrepancy across OMS, WMS, and Accounting.
Diagnose with evidence
Trace primitives return rejected candidates, override sources, recent changes. The LLM reasons over real data, with citations.
- "Why won't SKU-1234 publish to ecom-eu?"
- "Why is this price not used for customer X?"
- "What changed on this order in the last 24h?"
Edit by natural language
"Set material to cotton on SKU-1234" becomes a preview card. The args are shown. Nothing changes until you confirm.
- Bulk updates: "tag these 14 SKUs as clearance"
- Status flips: "mark these 6 returns as approved"
- Drafts: "create a draft promotion for next Tuesday"
Orchestrate across modules
The orchestrator calls tools across OMS, WMS, ACC, CRM in one turn. Users don't need to know which module owns what.
- "Why is Order-10 short by 5 units?"
- "Show me everything pending the close of Period-9"
- "Which active promos still need approval?"
400+ tools. 8 modules. One MCP transport.
Each module ships its own MCP catalog. The LLM can only do what its tools can do — that's the security boundary, not the prompt.
Per-tenant scope · per-user RBAC tier · per-module catalog · 20s per-tool timeout · prompt-injection delimiters · audit linkage on every write.
Same tools, your own MCP client.
The in-app web terminal is just one MCP client. Connect Claude Desktop, your own scripts, or anything that speaks Model Context Protocol — with the same auth and the same audit linkage.
No black box, no lock-in.
Every tool is reachable from outside the app. Customers can automate against the same catalog the in-app agent uses — from a CI job, a scheduled report, or another LLM.
- · Same auth: tenant-scoped, RBAC-tier-aware
- · Same writes-as-proposals contract
- · Same audit linkage to ChangeHistoryEntity
# Connect Claude Desktop to your Kubrik tenant { "mcpServers": { "kubrik": { "url": "https://api.kubrik.io/mcp/sse", "headers": { "Authorization": "Bearer $KUBRIK_TOKEN", "X-Tenant-ID": "$TENANT_ID" } } } }
Available in every Kubrik deployment.
Kubrik Copilot ships with Kubrik ERP, Kubrik POS, and Kubrik Commerce. No separate plan to buy, no add-on to enable — just an embedded AI layer waiting for the first question.