Security and control

AI should know exactly what it is allowed to do.

Reforception is designed around explicit connections, explicit model choice, bounded execution, and visible authority. Control is part of the workflow, not a warning page you read after something goes wrong.

Design boundaries

Access should be understandable before an agent starts.

01

Connections are explicit

Services, providers, repositories, and machines should appear because you connected them. Availability is visible instead of implied.

02

Model choice stays visible

A selected provider or model should not be silently replaced with another provider, worker, or billing path just to keep a task moving.

03

Execution is permission-aware

Reading context, working in an isolated workspace, running checks, publishing, and deployment are different kinds of authority. One should not automatically imply another.

04

Local and remote are labeled

A task running on a machine you control is a different boundary from a task running on a remote host. Reforception is designed to keep that distinction visible.

Authority is granular

“Can work” should not mean “can do anything.”

Different workflows can grant different ceilings. A worker may be able to inspect a repository and edit an isolated copy while still having no authority to publish, merge, deploy, or reach unrelated host resources.

Read repository

Inspect the bound project context.

Allowed

Work in isolated workspace

Make bounded changes away from published source.

Scoped

Run approved checks

Execute only the command kinds the workflow grants.

Limited

Publish / merge / deploy

Requires separate authority and is not implied by repair.

Not granted

Credentials and state

Public pages should never be the secret store.

Application credentials

Provider and service credentials belong to the authenticated application and its server-side connection paths, not this marketing website.

Truthful blocked states

Missing permission, unavailable capability, stale state, or an unconfirmed Stop should surface as blocked or uncertain rather than being presented as success.

Separated execution

Isolated workspaces and connected machines are different execution environments. Reforception is designed to preserve those boundaries rather than flatten them into one unrestricted shell.

Visible evidence

Changes, checks, worker state, and remaining decisions should be inspectable so completion does not depend on trusting a single model's confidence.

What this page does not claim

Design intent is not a certification.

  • No compliance certification is stated or implied.
  • This page is not a penetration test, warranty, or promise that misuse is impossible.
  • Reforception is in private beta; some boundaries and implementation details can still change as the product is completed and tested.
  • Features still under development are labeled as such elsewhere on the site rather than presented as shipped controls.