Skip to content

FAQ

A harness and a lab for intelligence you own. The harness is a desktop app with a browser, terminal, file explorer and models in one window. The lab is what it learns from: your memories, turned into training data and then into adapters. It runs on your hardware and keeps everything it learns there. For the longer answer, see What is Kaba? and Why Kaba?.

kaba is the desktop client. kabactl is the engine it starts: one binary that stores data, runs models and connects your devices. kabactl also runs alone on a server. See internals.

A lot. Kaba Personal is all local: you hold the account, the keys, the policies and the settings. With Kaba Enterprise and fleet management, sign-in, policy, enrolment and updates are controlled centrally by your organisation, which can also revoke keys and data remotely. See what fleet control changes.

Open source is pending. The source is available to read and build, and the client’s repository carries a GPLv3 No-AI Amendment licence (GPLv3-NAI).

[todo: confirm the licence statement and the public source URL. kaba.ai says open source; the kabactl page says proprietary; the kabactl crate declares GPLv3.]

In Kaba Personal you create a local account on your device. There is no sign-up with Kaba Labs and no account server. See accounts. On a fleet-managed device under Kaba Enterprise, sign-in is centralized and your organisation controls the account.

Not in Kaba Personal. Your password unlocks the key that encrypts your memories and saved logins, and nobody else has a copy. See security. On a fleet-managed device, ask your administrator.

Yes. Enable AI is off by default. Without it you have the tiling browser, privacy layer, terminal, file explorer and mesh. No model is downloaded and no model process starts.

Yes: Settings → Features → Enable Memory. With it off, nothing is captured automatically and the memory menus are hidden. You can also keep memory on and exclude specific sites with the Memory Ignore List.

Kaba Personal has no account server and sends no analytics. Your memories, prompts, outputs and adapters stay on your devices and move only between devices in your own cluster. The specific cases where something does go out, such as DNS queries and model downloads, are listed under when data leaves your device.

Is Ghost Mode the same as a private window?

Section titled “Is Ghost Mode the same as a private window?”

It is stricter and per pane. A Ghost pane has its own in-memory session, records no history, makes no memories, and is erased on close. It does not hide you from the sites you visit; combine it with Tor for that. See Ghost Mode.

A site says I’m a bot, or won’t let me sign in.

Section titled “A site says I’m a bot, or won’t let me sign in.”

Open the site panel for that site (click the site pill in the Kaba bar). Kaba pauses fingerprint protection automatically for sites that serve bot checks; if the check still fails, turn fingerprint protection off for that site. If part of the page is missing, allow third-party cookies for that site.

Check whether Lockdown is on globally (Settings → Security & Privacy). Lockdown turns JavaScript off everywhere. Turn it off there and use it per site instead.

Two sizes of Gemma are supported out of the box, e2b and e4b. You can import any GGUF model that has a chat template. See machine learning.

No for browsing, memory and the mesh. Inference runs on CPU if there is no GPU, more slowly. Training an adapter locally needs a GPU; without one, train on a peer that has one.

On the device your policy names: this one, or a peer in your cluster. Not in a cloud.

Only as far as your policy allows. By default it proposes commands and you click Run. File and code work happens in a container with no network. See security.

A small set of weights trained on top of a base model, here on your own memories. Adapters are files you own and can move between your devices. See models & training.

Create a ticket on one (Settings → Devices → Invite a device) and paste it on the other. See cluster & mesh.

No. Peers connect outbound and use a relay when they cannot reach each other directly. Traffic is end-to-end encrypted, so the relay cannot read it.

Yes. Set relays under cluster_options.

Yes. Linux is the first-class platform, and builds exist for macOS and Windows. What works where is under platform support.

A personal box running Kaba. See What is Kaba? and kaba-enclave.

See mobile.

The boot screen says the engine did not start.

Section titled “The boot screen says the engine did not start.”

Use Restart kabactl on that screen. If it persists, run kabactl doctor in a terminal; it checks the config, certificates, databases and ports. A common cause is another process holding port 28832.

A page will not load, but works in another browser.

Section titled “A page will not load, but works in another browser.”

Settings → Information → Troubleshoot → Probe. Paste the address; Kaba runs the load step by step through your session and reports where it fails.

Check Settings → Information → Diagnostics to see which process. Memory Saver frees idle panes; lower its timeout. If a model is loaded and you are not using it, right-click the header and choose Unload System Model.

Graphics glitches or a blank window on Linux.

Section titled “Graphics glitches or a blank window on Linux.”

Start once with KABA_VULKAN=0, or force the display platform with KABA_OZONE_PLATFORM=x11. See arguments & environment.

That is intentional. Run it as your user, or install system-wide with sudo kabactl service install --system --systemd. See kabactl.

kabactl ask or train complains about the store.

Section titled “kabactl ask or train complains about the store.”

Local ask and train need exclusive access. Stop the server first (kabactl service stop, or quit the client), or use --peer / --node to run on another device.