SentinelAgent

Local-first computer use

Your desktop.
Your control.

An AI agent that observes your screen, proposes actions, and asks before it acts.

Get started Explore the workflow
A sculptural silver S and graphite cursor on a white glass platform, representing human control.

Run locallyBring a model that fits your machine.

Review the actionKeep the next step in your hands.

Inspect the codeFollow the implementation on GitHub.

A practical experiment
in computer use.

From a terminal agent to a portable desktop app. Built to make the boundaries visible, not hide them.

Windows desktop

See the proposal.
Approve the next move.

Choose an open window, describe a task, and review each proposed input. Single-step mode is on by default.

Explore the workflow
Sentinel Desktop showing a proposed drag action, Approve once, Decline, and Stop controls.
Actual interface, shown with a test-fixture action. Not a live drawing result.

Terminal workflows

Start with an instruction.

The CLI includes allowlisted app launching, browser tasks, and direct Word, Excel and PDF workflows. 

sentinel-agent "Open Chrome,
open YouTube, and search
hello world" --no-model
--live-desktop

Illustrative instruction. Keep it on one line in your terminal. Availability depends on installed apps and workflow dependencies.

Portable Windows app

Keep your setup
with your app.

The desktop build bundles Python. Move the executable and its SentinelData folder together.

A graphite application tile beside a frosted glass folder, illustrating the app and its data traveling together.
Sentinel Desktop.exeSentinelData/

Ollama, model weights and optional Tesseract are installed separately.

The control loop

Intelligence proposes.
You decide.

Explore how an instruction becomes a reviewed action. This walkthrough is illustrative and never controls your computer.

Example: select a shape in Paint

Start with what is actually there.

Accessibility data identifies available controls. A vision model can also receive a window-only screenshot, with optional local OCR. 

Observation
  source       selected window
  controls     accessibility tree
  image        optional screenshot
  identifiers  fresh for this observation

The agent does not receive unrestricted access to every window.

Under the hood

Electron + React interface restricted IPC bridge Python agent model adapter + guarded execution

Choose the model.
Keep the control.

Profiles separate model configuration from execution policy. Changing providers does not remove the action checks.

Ollama

Use a local server. Browse to the executable, discover installed models, and test your profile.

127.0.0.1:11434

Compatible local servers

Connect LM Studio, llama.cpp, or another OpenAI-compatible endpoint using its server URL and model ID.

Cloud APIs, by choice

OpenAI-compatible, Anthropic and Gemini profiles are available. Sending observations to the cloud requires explicit task consent.

Built to be inspected.

A research project with measured checks and clear limitations. Not a claim of reliable, unrestricted autonomy.

Know the boundaries

  • Native control is experimental and targets one selected, permitted window.
  • Stop requests cancellation. It cannot undo input already sent.
  • Cloud consent and egress checks are not OS-wide network isolation.
  • Passing unit tests is not proof of success across arbitrary apps.

Take a closer look.

Run the desktop interface from source, or build the portable Windows executable yourself.

Run from source

Windows, Python 3.11+, Node.js 22.12+ and Git. Set up a local model server separately.

Read setup guide
PowerShell 
git clone --branch desktop-application https://github.com/IntrosWorld/SentinelAgent.git
cd SentinelAgent
python -m pip install -e ".[desktop,documents,dev]"
cd desktop
npm ci
npm start

Portable builds are unsigned previews. No hosted binary download is linked here; follow the README to build one.

A few things to know.

Does it work completely offline?

Local-model inference can run offline once the runtime and model weights are installed. Tasks that use websites still require internet access. Cloud providers are optional.

Can it control any app?

No universal reliability is claimed. Native actions use accessibility information and optional vision within a selected window. Elevated, sensitive, hidden or changed windows can be refused.

Where are the Word, Excel and PDF features?

Use the desktop conversation to create Word documents, Excel workbooks and PDFs. Append Word paragraphs or PDF pages, or set Excel cells. Import a copy of an existing document to edit it, review the proposed change, then open the verified output. These local commands need no model server.

What travels with the portable build?

Copy the executable and its SentinelData folder together while the app is closed. Python is bundled. Ollama, models and optional OCR are separate. Windows-encrypted API keys usually need re-entering on another account or PC.

Which workflows have been verified?

Real Electron integration tests create and edit Word, Excel and PDF files through the conversation, approvals and Python runtime, including backups and declined writes. Native window control uses accessibility and optional vision; its success depends on the application and model. An unlocked Windows desktop is required for live input.