GitHub - jrecyclebin/jabsco: Computer-use agent that works over RDP.

This is an RDP client that can be used to control a remote desktop using Claude. I chose RDP because you can use it to connect to Hyper-V virtual machines - even before an OS is installed.

So this also has Hyper-V specific capabilities - the ability to list, create and change VM settings remotely - which are given to the agent so that it can build VMs from scratch, access them over RDP and switch between them.

I also like remote desktop as the agent interface because it can't do anything outside of that connection - the virtual machine or remote user account is its sandbox.

Features

• GUI for chatting with the agent - also a working RDP client. (The biggest lift in this project was using the FreeRDP bindings directly to allow both agent and user access to the same session and have it be cross-platform.)

• You can have skills and commands, to bottle up prompts you want to reuse.

• Screenshots are taken automatically - and only one is kept in the conversation at a time - yeah that kills caching but it also keeps the context clean. (Screenshots can easily take 20k tokens each.)

• Observe mode: the agent locks the RDP client while it works, then releases control back to you when it's done. (Manual mode allows you to interfere at any time.)

• CLI mode. Run: jabsco --help for options. Use this to fire off single prompts.

• Works on Windows, Linux, and macOS. (I developed on Linux but tested on all three.)

Quick Start

Download a release - it's just a zip file with Jabsco.exe inside.

You'll need to put your Anthropic API key in the config file (see below) or set it as an environment variable (ANTHROPIC_API_KEY) before running it.

• macOS: Install the freerdp package with Homebrew: brew install freerdp. Requires FreeRDP 3.0 or later.
• Linux: Install FreeRDP with your package manager. On Arch, the freerdp package is sufficient.

There may be issues running on odd architectures - like Windows or Linux on ARM. Jabsco's custom bindings aren't quite as portable as I'd like.

Configuration

Jabsco reads config.toml from its config directory on startup:

• Linux/macOS: ~/.config/jabsco/config.toml
• Windows: %LOCALAPPDATA%\Jabsco\config.toml

If the file doesn't exist it is created automatically with commented-out examples. The directory can be overridden with the JABSCO_CONFIG_DIR environment variable.

# Your Anthropic API key.
# Priority: --api-key CLI flag > this value > ANTHROPIC_API_KEY env var
anthropic_api_key = "sk-ant-..."

[agent]
# Override the system prompt sent to the model on every request.
# system_prompt = """
# Your cursor position is shown as a red arrow.
# Additional instructions here.
# """

Hyper-V Support

Work is being done to allow connections to a Hyper-V host - or to any of its VMs. The connection to the host isn't an RDP connection - it's a dashboard of VMs that the agent can access. (This connection is only support on Windows, where there is access to Powershell remoting.)

This work is feature flagged - so you'll need to add an entry to your config.toml:

The connections to the VMs themseves are RDP connections - using Hyper-V's VMConnect functionality. On Linux and Mac, you'll need to drop the VM's ID in there on the Hyper-V connection tab.

All Hyper-V connections require you to use an actual Windows user's creds - and you'll also want to add that user to the Hyper-V Administrators group. (There's also likely work to do to get domain users working - you might want to create a local user on the machine that Jabsco can use.)

In order to list and createVMs, you'll have to set up PS remoting on the host:

> Enable-PSRemoting -Force

Then, you also need to install Hyper-V tools on the Windows machine you're running Jabsco on and get WinRM set up. (Run this in an elevated prompt.)

Set-Service WinRM -StartupType Automatic
Start-Service WinRM
Set-Item WSMan:\localhost\Client\TrustedHosts -Value 'HYPERV-HOST' -Concatenate -Force

To find out the VM IDs you'll need to use when connecting from Mac and Linux:

> Get-VM | Select Name, Id

Skills and commands

User-defined reusable prompts live under the config directory:

• Skills (skills/<name>/SKILL.md) — loaded by the model on demand via the load_skill tool. Useful for role or style instructions the agent can pull in when needed.
• Commands (commands/<name>.md) — invoked from the prompt with /name [args]. The prompt is replaced with the file's content; $ARGUMENTS is substituted with anything typed after the command name.

These are pretty limited compared to Claude Code but hey, it's something.

CLI Usage

The command-line interface is pretty basic for now, but you can run single prompts without opening the GUI:

jabsco run --host HOST-NAME --prompt "Open Notepad and type the first page of Oliver Twist"

If you previously logged into the machine at HOST-NAME using the GUI and saved the credentials, the CLI will use those.

A stream of NDJSON events will go to stdout.

Building Jabsco

Install mise.

Install .NET 10: mise use dotnet@10.

Then: mise build.