Private local agent · active

SEBAS

Private local assistant architecture built around explicit IPC, approvals and bounded system tools.

Verified Project State · 16.1

Verified Project State

SEBAS

intelligence-v16:sebas

fresh

Local assistant hardening, explicit approvals, policy budgets, inspectable system execution and resilient operator tooling.

Basis: 2026-08-18 · 0d · threshold 21d

Open Verified State Ledger →

Project Intelligence 3 · 2026-08-18

Living state

Local assistant hardening, explicit approvals, policy budgets, inspectable system execution and resilient operator tooling.

hardening

What changed

Core operator, memory, HTTP, file, Git and approval surfaces are established; the next public-safe story is hardening rather than remote control.

Updated 2026-08-18 · 21d freshness window

What’s next

  • Complete hardening
  • Expand policy verification
  • Improve approval UX
  • Keep execution local

Public dependencies

  • Yggdrasil · governance_peer — Shares local authority and explicit-approval patterns.
phase:hardeninglocal assistantapproval registrysystem policyinspectable tools

Overview

Problem

A local assistant becomes dangerous when language-model intent and system execution share an implicit authority boundary.

Design

Execution is separated behind explicit IPC contracts, policy checks and human approval surfaces for sensitive actions.

Outcome

An assistant model that treats system authority as a separate controlled capability rather than an automatic consequence of chat.

Verification

The public surface documents the authority model only.

Local agentIPCApprovalsSystem toolsAI agentsLocal automationSecurity

Architecture

  1. local assistant
  2. IPC boundary
  3. approval registry
  4. system policy
  5. safe executor
  6. tool surfaces
  7. memory window
Public architecture diagram for SEBAS

Components

component

IPC boundary

Local message transport between UI and assistant core.

component

System policy

Allow/deny and approval constraints around execution.

component

Approval registry

Durable pending decisions before sensitive operations.

component

Tool surfaces

Explicit local capabilities exposed under policy.

Decisions

active

Require explicit approval for sensitive execution

A local assistant should not treat model intent as authority to perform sensitive system actions.

Timeline

  1. Local IPC assistantmilestone · active
  2. Hardeningmilestone · ongoing
  3. Authority and approval architecture is public; IPC, commands and local memory remain private — public_boundaryevidence · 2026-08-11

Experiments

No additional public experiment records.

Releases

No project-specific public release record.

Evidence

public_boundary

Authority and approval architecture is public; IPC, commands and local memory remain private

Authority and approval architecture is public; IPC, commands and local memory remain private

source-contract · checked 2026-08-11 · fresh until 2027-08-11

Boundaries

No IPC endpoint, workstation control, command execution or local memory is exposed by the public website.