Skip to content
iMpact
In development

ORIONPersonal intelligence, reimagined.

ORION is iMpact’s flagship product, currently in development: a computer assistant designed to use permitted context, plan visibly, act within set limits, and return work for review.

Request early access Visit Labs

Current product scope

Capabilities we are building now.

  • 01

    Visual context

    We are developing a permission-bound way to interpret the interface and relevant task context.

  • 02

    Controlled action

    The product is being designed around visible steps, explicit permissions, interruptibility, and reversibility where possible.

  • 03

    Model routing

    A provider-agnostic routing foundation is in development around capability, latency, privacy, and cost.

  • 04

    Accessible paths

    Keyboard, voice, and assistive input are design goals being considered alongside the visual interface.

Architecture

Four layers, one accountable loop.

Select a stage to explore the accountable workflow we are building for ORION.

01 / 05

Understand

The intended flow reads the window, selection, and context you allow, then restates the task so it can be corrected before anything runs.

  1. 01

    Context layer

    Designed to read the active window, selection, and task history you permit, then build a short-lived picture of the work.

  2. 02

    Reasoning and routing

    Being developed to turn a request into a plan, then match each step to a model by capability, latency, privacy, and cost.

  3. 03

    Action layer

    Intended to run only permitted tools, one visible step at a time, with an immediate stop control.

  4. 04

    Review and record

    The design target is to return the result with its rationale and keep a readable history you can inspect, export, or delete.

Privacy and control

Boundaries are part of the product.

Local first where it counts
Sensitive work should stay on your machine. Where a request must leave the device, ORION says so before it happens.
Permission, not assumption
Access to files, apps, and the network is granted per capability and revocable at any time, without hunting through settings.
No silent training
Your content is not used to train models. There is no hidden opt-in and no quiet change of terms.
Accessible by design
Keyboard, screen-reader, and voice paths are built alongside the visual interface, not retrofitted after launch.

Direction

Current work, design targets, and open questions.

Nothing on this page describes a shipped, generally available feature.

Building now
  • Visual context engine
  • Permission-aware action layer
  • Multi-provider routing foundation
Design target
  • Keyboard and voice workflows
  • Readable activity history
  • Granular privacy controls
Open question
  • How far on-device models can carry daily work
  • Personal workflow memory without surveillance
  • A safe extension interface for developers
Request early access