How it works

One cycle, fully visible

Orbeex separates what you ask for from what actually runs. You approve the difference.

Ask

You state a goal in plain language, or use the regular tools and menus. Nothing runs yet.

Draft a plan

Orbeex turns the request into a typed set of operations. The AI proposes; it never touches data directly.

Validate

The plan is type-checked, policy-checked, and cost-estimated before anything is offered to you.

Approve

Safe steps can proceed; additive and destructive steps require explicit confirmation, with a preview.

Execute

The engine runs the plan locally. Each step becomes an operation in the history.

Review or undo

Every operation is recorded. Reproduce a result, or undo it, at any time.

The core idea

Everything is an operation

Every action — from you or from the AI — becomes a node in an append-only graph. That single rule gives Orbeex undo, audit, reproducibility, and scripting for free.

The operation graph

Operations form a directed graph: each one depends on the results of the ones before it. Change a step, and only the steps downstream recompute. The graph is the record of how a result was produced.

ask    →  buffer(river, 1km)
       →  join(schools, buffer)
       →  dissolve(joined)
       →  count()          ← result

The safety gate

Between a proposed plan and execution sits a validation layer. It confirms the plan is well-typed, allowed by policy, and within your stated budget — before you are ever asked to approve it.

Type and schema checks
Policy and risk-tier checks
Cost estimate per session
Dry run for destructive steps