agent-seat

Gives an AI agent its own Windows desktop session: screenshots, UI text, mouse/keyboard control and a file hand-off folder through a local CLI or stdio MCP server, while the owner keeps using their own desktop. Windows 11, experimental preview.

GitHub

Documentation

Give your AI a Windows desktop. Keep using yours.

English · 한국어 · 简体中文

Download · Install & recovery · CLI / MCP · Security

An AI writes a note and runs a calculation inside its own Windows session.

_{Captured from the AI seat itself through the CLI. The owner's desktop is not involved.}

Why agent-seat?

Your agent needs a place to open apps, read screens and carry out desktop tasks. You still need a computer you can use.

agent-seat creates a separate local Windows account and desktop session for your AI. Your agent works in that session while you keep using your own. Open the local viewer to see its progress, pause input or take over with your mouse and keyboard.

Use it to give an existing AI agent a desktop workspace, explore Windows app automation, or supervise a task without keeping the AI's screen in the foreground. The project provides the desktop and control tools; bring your own image-capable agent or model client.

Experimental preview · 0.9.3. The installer uses an unsupported TermWrap modification to enable simultaneous sessions on Windows clients. Windows updates can affect compatibility. Fresh-machine installation across Windows builds and restart/re-login are still unvalidated. Review the validation record before installing.

What you can do

Capability	What it gives you
🖥️ Dedicated desktop	A standard Windows account and its own interactive session
👀 Local viewer	See the AI's desktop in a browser; starts read-only
🖱️ Human takeover	Click, double-click, right-click, drag, scroll, send keys and IME text
🧰 CLI + optional MCP	Connect your existing agent through either interface
📝 Selective observations	UI text, unchanged-frame detection and changed-region crops
⏸️ Owner controls	Pause, resume and stop input from the local dashboard
🌐 Three UI languages	English, Korean and Simplified Chinese with a saved language choice

The release includes the service, CLI, RDP anchor, input helper and .NET runtime.

How it fits together

flowchart LR
    You[You] --> Own[Your Windows desktop]
    Agent[Your AI agent] --> Tools[CLI or MCP]
    Tools --> Service[agent-seat local service]
    Service --> Seat[Dedicated Windows desktop]
    You --> Viewer[Local browser viewer]
    Viewer <--> Service

The banner is a conceptual illustration. Both desktops share one Windows host. Separate accounts and sessions are not a VM or security sandbox.

Quick start

Double-click installation

Download Install-AgentSeat.cmd and double-click it. It downloads the executable release ZIP, verifies its SHA-256 checksum and GitHub asset digest, extracts it and starts setup. No manual extraction or developer tools are needed.

Review the installation plan, type INSTALL to accept the unsupported Windows client modification, then approve Windows administrator access using the same owner account. Setup creates the seat and opens its screen read-only.

One PowerShell command

Open 64-bit Windows PowerShell under your ordinary installing owner account and run:

& ([scriptblock]::Create((irm 'https://raw.githubusercontent.com/aksmfosef11/agent-seat/v0.9.3/Get-AgentSeat.ps1')))

This downloads and executes the versioned bootstrap script from this repository. The same installation review and Windows administrator prompt follow. The bootstrap checks the ZIP before running its installer.

To let your AI install it, give it the AI-run installation guide. It lists the administrator-access and consent steps the AI must hand back to you.

Offline ZIP installation

Download agent-seat-0.9.3-win-x64.zip and SHA256SUMS.txt from Releases, compare the hash, then extract the ZIP and double-click its Install-AgentSeat.cmd. This path uses the local package without another download.

Get-FileHash .\agent-seat-0.9.3-win-x64.zip -Algorithm SHA256

GitHub's automatic source archives do not contain built applications. End users do not need Visual Studio, Node.js, Git or a .NET SDK.

Setup reuses an existing seat and opens its viewer; it does not reset accounts or upgrade existing application binaries. See installation and recovery for options, download-only mode and interrupted installations.

After setup, View-Seat.cmd opens the screen again. Stop the AI task before enabling Manual control. Use the text box for IME text and the shortcut buttons for combinations intercepted by the browser.

Requirements

Item	Preview support
Host	Windows 11 Pro / Enterprise / Education, x64
Installation	Administrator access from the interactive owner's account
AI client	Your own agent that can read images and call CLI or MCP tools
Runtime	Included in the release ZIP (.NET 8.0.31)
Unsupported installers	Windows Home, ARM64 and Windows Server/RDS

The anchor runs under the installing owner. After logging out and back in, start the seat again with computer start. Review recovery and current limitations for interrupted installations and Windows updates.

Connect your agent

CLI

MCP is optional. The CLI reads the protected token for the installing owner:

$cli = "$env:ProgramFiles\agent-seat\cli\agent-seat.exe"
& $cli computer start --seat agent
& $cli computer begin --seat agent
& $cli computer guide

Start a new observation context with begin, open its returned screenshot, then reuse its context ID for subsequent observations and actions. Input can be batched; observe the result before retrying a failed action. See the English CLI workflow.

MCP

Register the same executable with arguments computer mcp --seat agent in your MCP client. Four tools are available: seat_status, seat_start, seat_observe and seat_act. Images can be returned directly in tool results. See MCP setup.

Token usage: UI text, image reuse, crops and batches can reduce repeated observations. Savings depend on the model and task; no percentage is promised. The human viewer itself sends no frames to a model and makes no model API calls.

Languages

The dashboard and viewer support English · 한국어 · 中文（简体）. The first visit follows your browser's preferred languages; unsupported languages fall back to English. Use the language selector in the header or viewer to change it. Your choice is saved in that browser without reloading the page or clearing drafted text.

This changes agent-seat's interface, not the Windows language or the apps inside the seat. CLI/API machine messages and original diagnostics stay in English for compatibility. Unknown provider diagnostics remain available rather than being guessed or hidden.

Build and contribute

Developers need Git, .NET 8 SDK, Visual Studio/Build Tools with Desktop development with C++ and the Windows SDK. Node.js is needed for UI tests.

git clone https://github.com/aksmfosef11/agent-seat.git
cd agent-seat
.\scripts\Build-TermWrap.ps1
dotnet test .\AgentSeat.sln -c Release
npm test
powershell -NoProfile -File tests/installer/bootstrap.tests.ps1
.\scripts\Package-Release.ps1

Commit source changes before packaging; the release records its source commit. Generated binaries and local credentials stay out of Git. See releasing and localization.

Reports with a Windows edition/build and clear reproduction steps help most. Remove tokens, credentials and personal screenshots before opening an issue. Fresh-install and re-login testing are especially useful before a stable release.

Documentation

Guide	English	한국어	简体中文
Project overview	README	README	README
Install & recovery	Guide	안내	指南
CLI workflow	Guide	안내	指南
Optional MCP	Guide	안내	指南
AI-run install	Guide	안내	指南

Security · Validation · Architecture

License

MIT for this project's source. Dependencies retain their original licenses; see third-party notices.