$ kaba --help
Attention Is All You Need,
Train Yours.
One binary. Your devices become a private mesh that runs, trains, and remembers. Free and open source. Local and encrypted by default. Runs anywhere. Yours.
["a harness. a browser. a terminal. a file manager. one surface. your own lab."]
Using models is now the norm. With Kaba, build datasets, train with RL, and turn telemetry into fuel. Be your own lab.
Full stack, end-to-end context — the reason the models get good.
No other tool, app, or model has your full context — just the prompts, or just the data within that app. Kaba's client encapsulates your flow across the entire stack, observing the ground truth of how, when, and why you work. Every observation stays local, encrypted at rest.
[observe everything, leak nothing]
["straight from the wire — these are the real protocols"]
// THE MESH — peers, streams, tokens
Every stream is an encrypted QUIC connection between peers you own — addressed by key-derived .kaba names over iroh, no DNS, no CA. Prompt goes out on kaba/infer/v1, tokens stream back. Anything that leaves the mesh can exit over Tor — and adapter selection is the peer's policy, the requester never chooses.
// THE LEARNING LOOP — observe → memory → RL → adapter
Memories are written autonomously from observed context, encrypted before rest. RL runs over memories and trajectories; the output is a LoRA adapter that is yours — sell, buy, trade, breed.
toolloop
// THE TOOL LOOP
01 prompt → plan
02 tool call → gVisor container
03 observe → checkpoint
04 subagents → peer GPUs
05 hooks gate every edit
["everything local, everything encrypted, everything yours"]
cluster::gossip
Auto-clustering
Your devices discover each other and form a mesh over iroh — gossip, heartbeat, join tickets. No config, no cloud broker, no open ports.
kaba_engine
One harness LAB, all models
A single model harness LAB with backend selection per device — GPU or CPU, any local model. Quantize, cache, and probe VRAM automatically.
kaba/train/v1
LoRA build pipeline
Build adapters from your proprietary data and workflows. Train locally, or push the job to a clustered peer for more resources. Import, manifest, and version every adapter as you scale.
memstore + memcrypt
Autonomous memories
Kaba observes what you actually do and writes memories on its own — verifiable, private, encrypted at rest before they ever touch disk.
rl::memories
RL on your memories
Reinforcement learning runs over your memory graph. The model improves from what it observed about you, not from someone else’s dataset.
trajectory + telemetry
Trajectories feed training
Every run is checkpointed as a trajectory. Rewind any run, replay it, or train on it. All telemetry is fuel for your hill climb.
toolloop
Full tool loop, any local LLM
The canonical tool loop runs in the daemon: every exec sandboxed in a gVisor container, hooks, subagents fanning out to peers. Any local model becomes an operator.
helm install kaba
Orchestration-ready
A harness LAB that scales with your platform across Kubernetes, Nomad, and any container cluster you run. Carve the mesh per user, per business unit, per org: same binary, scaled on your systems.
tor + .kaba
Anonymous by design
Tor as a first-class transport, key-derived .kaba names over iroh — no DNS, no CA. Plus Ghost Mode: incognito on steroids, zero trace on disk. Both toggle per-pane.
policy::gate
Local-only, policy-driven
RAG, models, adapters, settings, and memories are all encrypted at rest and gated by policies you write. Nothing leaves unless a policy says so.
run::anywhere
Platform-agnostic, default local
Linux as first-class citizen with native macOS and Windows support. Runs on desktops, servers, headless IoT boards, or ESP32 microcontrollers. Same binary, any edge. No cloud dependency, no vendor lock-in.
kaba://*.kaba
Sovereign service exposure
Expose anything under a .kaba domain. Publish APIs, dashboards, data, or content behind a key-derived .kaba name. Access is either public or restricted to your chosen peers. Iroh/QUIC direct connect means no DNS lookup, no CA, no open ports.
["your data, your alpha, no compromises"]
[a] CONSUMER
A lab of your own
Your laptop, desktop, and home lab become one private AI. It learns from what you do, remembers for you, and is aligned to you. Train on your environment as you learn, work, and play as usual.
[learn, work, play]
[b] SMB
AI without the subscription
Turn prior investments in hardware and IT into a shared grid for inference and training. Predictable cost, and no customer data leaves the office — packaged with frontier capability.
[own, not rent]
[c] ENTERPRISE
Verifiable by construction
Policy-driven everything: who can infer, what can train, where data rests. Trajectories give you a full audit trail — GDPR and CCPA answers you can prove. Supercharge your people, product, and process without compromise.
[audit, verify, report]
[d] GOVERNMENT
Air-gapped intelligence
Runs fully disconnected. Open source, open hardware, encryption at rest, no external dependency to certify. Sovereignty is the default, not a SKU.
[sovereign, air-gapped]
["a remote control for the whole mesh"]
All the benefits of the client in an on-the-go package. Never walk around with your laptop open again.
chatFull bi-directional conversations with any model on the cluster — pick a peer, pick an expert, talk.
remoteA remote for every peer in the mesh — see what's running, start and stop jobs from anywhere.
manageDevices and policies in your pocket — approve peers, edit rules, watch the mesh react live.
["if the client is the hero, the CLI is the stunt double"]
Everything the client does, scriptable.
Join clusters, stream inference, push training jobs, inspect memories, edit policies — from any shell, over the same encrypted streams. Pipe it, cron it, put it in CI.
$ kabactl --help
cluster join, leave, status, ticket
infer stream tokens from any peer
train lora build · push · import
mem ls, show, forget — your memories
policy edit the rules the mesh enforces
fs browse and stream peer files
$
stdin | kabactl
Composable
Reads stdin, writes JSON. Every command pipes into the next. Your mesh speaks Unix.
--peer anywhere
Peer-addressed
Target any device you own by name. LAN, tailnet, or across the world — the stream finds it and connects you directly.
same policy::gate
Same rules apply
The CLI goes through the identical policy gate as the client. No backdoor, no service account, no exceptions.
["you did the work. you own the expert. prove it and profit."]
Every adapter you train is an expert. Experts are assets.
The LoRAs you build, the datasets you curate, the tools you design — list them, price them, sell or lease them. And because experts are composable, buyers don't just run yours: they breed it with their own — merging adapters into offspring with genuinely new capability. Unique thought, compounding.
[sell · lease · breed · profit]
market::adapters
Experts for sale or lease
List a LoRA with its manifest and eval trail. Sell it outright or lease it by the epoch. Buyers verify provenance before a single token streams.
market::datasets
Datasets, policy-scrubbed
Curated training sets are built from your work and released only through the policy gate. What you sell is exactly what you chose to share.
market::tools
Tools for the loop
Ship the tools you design for your experts — sandboxed, gVisor-contained, ready to drop into anyone's tool loop.
expert::breed
Breed for new capability
Combine experts — yours with bought, bought with leased. Merged adapters inherit from both lines; the offspring is a new asset you can list.
["is there a lib for this?" — the house question]
Your Hardware. Your Model.
Start Training Your Attention.
$ curl -fsSL https://kaba.ai/install.sh | sh
>- DOWNLOAD KABA FOR {{ osName }}
free · open source · one binary
COMING SOON
[click anywhere to close]
[click anywhere to close]