Updated October 18, 2026.
Screenshot of the official Kaba homepage, captured October 18, 2026. These are vendor marketing statements; current download availability and open-source status remain unverified, as explained below. No software was installed.
Kaba AI is Kaba Labs’ desktop workspace for browsing, local AI, searchable memories and work across your own devices. Its current documentation describes two parts: Kaba Console, the desktop client, and kabactl, the engine for storage, models and peer connections. The documentation covers Console 0.146 and kabactl 0.87.
Availability checked October 18, 2026: the official website and documentation are live, but we could not verify a working public installation route. The advertised download and installer addresses returned website HTML during our checks, while the linked core source and release addresses returned 404 responses. Treat the features below as documented capabilities, rather than results from a hands-on product test. Confirm a current release with Kaba before planning a deployment.
What does Kaba AI do?
Kaba brings a browser, terminal, file explorer and AI tools into a window organized with tiled panes. Its local engine can also run without the desktop interface on another machine. That makes its scope broader than a conventional notes app: it combines a workspace, a memory store and tools for running and adapting models.
The official documentation distinguishes Kaba Personal, controlled by the individual, from Enterprise fleet management, controlled by an organization. This distinction matters when evaluating who can set policies or access activity.
Documented features
- Searchable memories: Hippocampus stores snapshots of pages and, when enabled, terminal sessions, file activity and project conversations. Its documented tools include semantic search, version history, annotations and training weights. This is more specific than a promise of perfect long-term recall.
- Local model tools: selected text can be summarized, explained, rewritten or checked with an on-device model. Projects provide a place for longer interactions. AI is off by default; the documentation says model downloads require a user action.
- Model and adapter management: users can import compatible GGUF models and choose a base model and LoRA adapter through a policy. A LoRA adapter is a smaller set of learned weights used with a base model.
- Training on selected memories: the documented training flow can filter records by date, domain, topic or keyword. Training needs a GPU locally or on a connected peer. An adapter is not a guarantee of more accurate answers.
- Device connections: a ticket-based cluster connects your machines for tasks such as remote inference, terminal access and file browsing. Policies determine which device runs the work and what tools it may use.
Sources: Hippocampus documentation, models and training, and settings and devices.
Pricing and open-source status
The Kaba website advertises Kaba Personal as free. We did not verify a paid Personal tier or a public Enterprise price list. Enterprise visitors are directed to contact Kaba. Hardware, storage and running costs remain part of evaluating a self-hosted setup.
The licensing information is inconsistent. Although the homepage uses the open-source label, the official FAQ says open source is pending. It names a GPLv3 No-AI Amendment license for the client and explicitly flags conflicting statements about the engine. The public source links we checked did not resolve to an accessible core repository. Verify the actual license and source availability for the version you obtain before assuming modification, redistribution or commercial-use rights.
Platforms and release availability
The platform guide describes Linux as the primary platform and lists macOS and Windows desktop packages. It also leaves minimum operating-system versions and some platform details unconfirmed. Inference can use the CPU; local adapter training requires a GPU. Published minimum and recommended RAM or VRAM requirements are incomplete.
The mobile documentation states that the companion app is outside the documented releases and its store listings are not live. Do not read the homepage’s mobile badges as confirmation that an iOS or Android app is available.
The Enterprise guide also separates existing building blocks from fleet-scale capabilities outside the documented release. Confirm the availability of centralized administration, approvals, fleet search and other required Enterprise features directly with the vendor.
Privacy: what local-first actually means
Kaba’s documentation describes local storage and inference, with optional work on your own peers. That still involves network activity: browsing reaches websites, DNS uses configured resolvers, and models, updates and filter lists come from external services. A mesh connection may use a relay. The privacy guide explains these cases.
Memory capture is on by default, even though AI is off. Review Settings → Features and the Memory Ignore List before browsing confidential material. Ghost panes are documented as excluding activity from history and memory, but that does not establish anonymity to websites or networks.
The security guide says memories and saved credentials are encrypted, while the database containing settings, history, frames and policies is not encrypted by Kaba. It recommends full-disk encryption. It also warns that kabactl 0.87’s API and proxy listen on all network interfaces, with ports 28832 and 28833 requiring firewall protection on exposed machines. These documented limits rule out blanket claims that all data is encrypted or that a default installation guarantees privacy.
A practical way to evaluate Kaba
- Verify access first. Obtain a working release, its requirements and its license through Kaba’s official channels. We have not validated the installation command shown on the website.
- Choose a small, non-sensitive task. For example, collect a few public research pages and check whether you can find the right passage later.
- Review capture controls. Decide whether automatic memory is useful, exclude sensitive sites, and inspect what a saved memory actually contains.
- Keep execution limited. The documented starting controls are command execution set to Ask and sandbox networking set to Off. Check which device and model a policy selects.
- Check answers against originals. Compare a summary or retrieved answer with its source page. Record missing context or incorrect statements before trusting the workflow with important work.
- Evaluate training separately. If you later train an adapter, use a curated sample and compare it with the base model on questions held out of training.
This is a suggested evaluation workflow based on the getting-started guide and the documentation linked above. We did not install Kaba or run these tests.
Who should consider it?
Kaba is worth investigating for technically comfortable users interested in local models, research memory and connecting their own computers. Its documented controls also make it relevant to developers evaluating self-hosted AI workflows. Public installation access, licensing clarity, incomplete requirements and release-specific feature availability should be resolved before relying on it for everyday or organizational work.
For a wider comparison, browse the GPTOnline AI tools directory and compare tools by verified availability, data handling and the task you need to complete.
