SMALL SOFTWARE. USEFUL INTELLIGENCE.INDEPENDENTLY MADE BY BRUKET
Lavett.

More useful work.
Less wasted effort.

H HOLD MY SANDBOX LINUX · V0.2.0

Your agent’s desktop.
Your computer, still yours.

Let the agent work the interface.
Keep getting on with your day.

hms lets Claude Code and Codex operate browsers and desktop apps in a graphical session of their own, without taking over your screen, mouse or keyboard.

Your existing AI client One place for GUI work

A separate workspace, a checked resultA technical schematic of an agent program, a separate graphical session containing browser and document windows, and a returned result.HMS / SESSION TOPOLOGYPLATE 02SEPARATE LOCAL SESSIONagent / program01 open02 read03 write04 verifybrowserdocument✓01 / PLAN02 / OPERATE03 / CHECKILLUSTRATIVE SCHEMATICLOCAL / INSPECTABLE
FIG. 02 / A SESSION OF ITS OWN
01 Your pointer stays yours02 Browser & desktop work03 Results checked locally

01 ONE TASK. SEVERAL APPS.

From “can you…”
to “here it is.”

Describe the job in your usual conversation. Your agent can read a page, work in an application and check the result through the same interface.

AN EXAMPLE REQUEST

“Read the event details and save a short briefing in a document.”

One agent. Its own session.
You keep your desktop.

hms / separate local session
○ ○ ○ Event page

THE DESIGN ASSEMBLY

A day for
making things.

Friday · 10:00–16:00
Workshop Hall

○ ○ ○ briefing.txt

EVENT BRIEFING

Design Assembly
Friday, 10:00–16:00
Location: Workshop Hall

Bring your project notes.

✓ Document saved & checked
Illustrative example · result shown01 → 02 → 03

Illustration of the workflow, not a live agent or a recording. This page does not open applications or create files.

02 FEWER ROUND TRIPS

Let code handle
the mechanics.

When the steps are known, hms can execute a complete program locally. It waits for real readiness events and checks effects, instead of asking a model what to do after every click.

01

A whole plan, in one go.

The agent writes the GUI steps together. The runtime carries them out and returns the requested observations.

02

Read what is actually there.

Page text and accessibility data offer shortcuts. Pixels and text recognition cover interfaces without that structure.

03

Explore when the way is unclear.

An optional navigator can work through an unfamiliar interface. The calling agent can observe, steer and stop it.

03 PUT IT TO WORK

A session of its own.
One install away.

The current Linux release is tested on CachyOS with KDE Plasma 6.8 on Wayland. Other distributions are untested. Check the requirements before installing.

Read the system requirements ↗
LINUX / HMS 0.2.0Available

Run in your terminal

curl -fsSL https://lavett.dev/hms/latest/install.sh | sh

Installs into its own Python environment and registers with Claude Code and Codex. The installer never uses sudo.

macOS is in testing and is not qualified for this release. Windows is not available.

04 KNOW YOUR WORKSPACE

A little separation.
A clearer boundary.

Your files & session

On Linux, the session hides your home and exposes other files read-only, with ~/hms-exchange as the shared writable folder. Services your account can use, such as Docker or libvirt, remain reachable and can change files with their own rights.

Understand the boundary ↗

Your AI & subscription

hms uses the CLI sign-in you already have. Your provider receives what the agent sends it. The navigator stops when the CLI reports a subscription limit; the step that reaches the limit can still be billed as extra usage.

Read the provider details ↗

FROM THE SAME WORKSHOP

Give the next task
a better starting point.

Explore Lavett’s approach to context, memory and making your AI budget go further.

Explore Lavett ↗