Kaba Stack

An introduction into your Kaba environment out of the box.

Client

The client (kaba) is the desktop app. It launches like an application but houses an entire world: a tiling browser, a terminal, a file explorer, a desktop of widgets, a memory store and a model harness, all in one window. Web traffic passes through its privacy layer, and anything a model does passes through a policy.

Part	Docs
Kaba bar, frames and panes	UI layout, Kaba bar, tiling windows
Omnibar	Omnibar
Terminal, file explorer, desktop	Terminal, file explorer, desktop
Memory	Hippocampus
Models and tools	Machine learning, projects & tool loop

Kabactl

kabactl is the engine the client starts. It is one binary that serves the local API and proxy, stores memories, runs model inference and training, and joins your devices into an encrypted cluster. On a server it runs by itself as a headless node.

Part	Docs
Install and run	Install & quick start, services
Commands and configuration	CLI reference, config file
Cluster	Cluster & mesh, protocols
Models	Models & training

The four layers

Kaba describes itself in four layers. Each maps to parts of the client and the engine.

Harness: browser, terminal, files, models

One place to run every model. Kaba does the work itself, in the same windows you do.

Part	Provided by	Docs
Frames and tiling panes	Client shell	UI layout, tiling windows
Browser with a privacy layer	Client	Privacy
Terminal, local or on a peer	Client and engine	Terminal
File explorer, local or on a peer	Client and engine	File explorer
Chat, tool loop, adapters	Client and engine	Projects & tool loop

Mesh: peers, VPN, exit nodes

Every device you own joins one encrypted private network. No accounts with a third party, no cloud broker, no open ports.

Part	Provided by	Docs
Invite, join, evict	Engine	Cluster & mesh
Named streams between peers	Engine	Protocols
Exit nodes	Engine proxy	Services
Tor, per pane or global	Engine and client	Privacy
Exposed services	Engine	Protocols

Policy: every step checked first

A policy says where inference runs, which model and adapter answer, which tools may be used, whether commands may execute, and whether the sandbox gets network. Policy ceilings always win: nothing lower down can re-enable what a policy removes.

Part	Provided by	Docs
Policy editor	Client settings	Machine learning
Tool overlays and flows	Toolbench	Projects & tool loop
Command allowlist and sandbox	Engine	Security
Portable policy bundles (.kabap)	Client	File formats

Lab: memory, datasets, training

Everything you see and do can become private, encrypted memory. Memory becomes training data, and training produces adapters that are yours.

Part	Provided by	Docs
Memory capture and search	Client and engine	Hippocampus
Learning data (trajectories)	Client and engine	Settings
Training, local or on a peer	Engine	Models & training
Models and adapters library	Client settings	Machine learning

The loop

Run. Observe. Train. Repeat. You work in the harness. What you choose to keep becomes memory. Training turns memory into an adapter. The next run uses it. The loop stays on your hardware, and each part of it can be switched off.