ARGUS

ARGUS is a local-first Windows assistant for one operator. It combines voice and typed commands, a Svelte desktop HUD, a FastAPI control plane, bounded agent workflows, desktop automation and defensive monitoring. It is in active development; backend policy controls permission to act.

Role
Design, engineering and security
Year
2026
Status
In active development
Code
Public repository

Current source release

Updated from the public repository on 23 October 2026. Source and documentation are linked below. The recorded screenshots and demo show an earlier build; they are not a live session. No new Windows runtime or security test result is claimed.

Desktop HUD
Svelte 5
Local API
FastAPI
Authorization levels
L0–L5
Source license
GPL-3.0+

Overview

Purpose

A practical assistant for a trusted Windows account. Models can propose work, but policy, authentication, confirmation and execution controls decide which actions may run. ARGUS is not a remote-control service, multi-tenant platform or replacement for endpoint protection.

Desktop and productivity

Registered skills cover applications, windows, files, system status, timers, reminders, schedules, clipboard and local notes. Browser, research, documents, email, phone and messaging paths depend on their installed integrations, credentials and policy. A module in the source is not a promise that it is enabled on every installation.

Agents and interface

The V2 HUD includes tasks, agents, modules, security, logs, tools, settings, notifications and a command surface. Agent work uses typed goals, budgets, bounded retries, evidence and verifiers. Plans and replanned alternatives are staged for owner approval; each step is authorized separately. The 3D city and office visualize state, not an additional permission path.

Active development · 2026 Source reviewed:

Case-study summary

Problem
Develop a Windows assistant with explicit separation between AI-generated proposals and permission to execute actions.
Architecture
Svelte 5 / TypeScript HUD → FastAPI REST and WebSocket service → routing, authorization, action-bound grants and registered skills; separate voice workers.
My contribution
Developing the assistant, V2 HUD, agent kernel, Windows skills, monitoring and backend security controls.
Tests
The case study references public source code and documentation. Current Windows runtime test results have not been independently verified.
Outcome
Public Windows source release with voice, desktop workflows, bounded agents, a 3D state view and documented setup.
Limitations
One trusted Windows operator; partial sampled monitoring; optional integrations require setup and may send data externally. Not endpoint protection or a signed installer.

Project evidence

Architecture

Service and HUD

Python with FastAPI/Uvicorn serves REST endpoints and WebSocket events over loopback. The V2 HUD uses Svelte 5, TypeScript, Vite and Three.js. Nonpublic REST routes and the WebSocket require a per-session token; allowed browser origins, body limits and rate limits constrain access.

Voice and models

Wake handling and speech activity detection feed speech-to-text, routing, response generation and speech synthesis. Speech workers are separated from the main service. Ollama provides the default local conversation and routing path; Whisper and Piper require their local models. Optional hosted models and web services have separate egress and data-scope controls. Local-first does not mean every configured feature works offline.

Command path

HUD or microphone → local service → intent rule or model route → named skill, action and target → policy and authentication → confirmation and action-bound grant where required → execution → audit, observation and response. Model output is input to this flow, not an execution credential.

Security

Authentication and grants

L0 covers low-risk responses; L1 requires unlocking; L2 requires fresh authentication; L3 adds explicit confirmation; L4 requires stronger multi-factor confirmation; L5 requires a hardware-backed gesture where configured and supported. Missing factors refuse protected actions. Short-lived grants bind approval to the canonical action and target and are checked immediately before execution.

Containment and recovery

Execution policy, declared skill permissions, sandboxing, network egress rules, integrity checks and audit-chain checks provide layers of control. Unknown network components are denied by default. Deterministic security state can move through NORMAL, DEGRADED, SAFE_MODE, LOCKDOWN and RECOVERY. These controls do not prove that every skill is isolated or safe.

Trust boundaries and limits

Voice recognition is not authentication; an ordinary RGB camera is not a reliable sole identity factor. Software already running as the same Windows user may read or change local configuration, tokens, code and audit material. Local integrity is tamper evidence, not a hardware trust root. Optional integrations can send permitted data to external providers.

Monitoring

Defensive observations

ARGUS samples selected Windows persistence, process, privacy, USB, network and LSASS-handle signals. Coverage depends on the collector, permissions and optional helper; it is not a replacement for endpoint protection.

Collector availability

An optional elevated LSASS helper has a separate IPC boundary. Some observations require Windows APIs or additional permissions. A collector that cannot run must be shown as unavailable or degraded. Short-lived activity may occur between samples, so no report proves that the host is free of compromise.

Detection engineering overview

Verification and setup

Evidence for this page

This page was checked against public source revision 83fbdca and its specification, architecture, feature catalog and security model. The repository does not publish the historical 109-suite test set cited by the earlier case study. Those totals and absolute test-proven claims have been removed from the current project summary. This update is a source review, not an independent security audit or Windows runtime test.

Source installation

Windows, Python, Node.js and npm are prerequisites; Ollama and voice models are needed for their local AI and speech features. Use a virtual environment, install dependencies, copy the example configuration, set the local PIN hash and build the V2 HUD. The source release excludes credentials, private runtime data, model weights, a signed installer and the optional external 3D office pack. Code-generated geometry remains available without that pack.

ARGUS AI City from above: an isometric city where every robot is an ARGUS agent
Skip the flight

ARGUS City

Every robot is a real ARGUS agent. Every light is backend state.

A team forms

“Why is my PC slow?” The Agent Manager staffs a team, and they head for Agent HQ.

Policy and auth gate

The agents decide what they want to do.

Policy and auth gate

Only the policy and authorization gate decides whether they may.

The only road to machine control. Every action stops here for policy, and for you.

Built by Sodik Tursunboev

ARGUS City

Designed, engineered and secured by one person, and still in active development.

ARGUS in 51 seconds

Recorded demonstration from an earlier ARGUS build. It illustrates the interface, agent teamwork and authorization concept. Use the current repository documentation for available features, setup and security limits.

  1. ARGUS

    A personal AI operating system: voice, local AI and security, on your own PC.

  2. The HUD

    Local AI, live telemetry and a voice you can talk to, all running on this machine.

  3. Security first

    Sealed files, a policy engine and a tamper-evident audit chain. The AI never decides what it is allowed to do.

  4. ARGUS City

    Every robot is a real ARGUS agent. Every light is backend state.

  5. A team forms

    “Why is my PC slow?” The Agent Manager staffs a team, and they head for Agent HQ.

  6. Agent HQ

    Planner, System, Diagnostics and Network work the task graph. A Log Analysis specialist is hired for the job.

  7. Verification Institute

    Nothing is accepted until the Verifier has checked the evidence.

  8. Cloud Embassy

    A cloud specialist researches here, outside the secure city. Cloud can think. Cloud cannot control.

  9. “Send everyone to the academy”

    One spoken order, and all ten agents walk to class.

  10. ARGUS Academy

    Lessons run one at a time on the local model, and a verifier grades every one.

  11. Policy and auth gate

    The only road to machine control. Every action stops here for policy, and for you.

  12. Built by Sodik Tursunboev

    Designed, engineered and secured by one person, and still in active development.

Next projectCustom SIEM + Sentinel BASA Python SIEM validated against real Atomic Red Team techniques.

Contact and socials

ARGUS source is public. See the repository for the current implementation, or send a message below for a walkthrough.