Kaba — Intelligence you own.

Intelligence
you own.

Runs on your hardware.
Learns from you.
Keeps everything it learns
and builds everything you need.

Free · open source · one binary
Kaba Labs — the harness and lab for intelligence you own
⤢ Fullscreen
Harness — browser, terminal, files, models Mesh — peers, VPN, exit nodes Policy — every step checked first Lab — memory, datasets, training

You are the best training data there is. Right now you're training someone else's model. Kaba keeps it yours.

Harness

One place to run every model.

Browser, terminal, files, and models in a single harness on your devices. Kaba does the work itself, in the same windows you do. Policy checks every step before it runs.

Kaba harness
Memory

It remembers what matters to you.

Everything you see and do becomes private, encrypted memory, curated into clean datasets automatically.

Kaba memories
Lab

Then it gets better at it.

Train expert models on your own context, on your own GPUs or a peer’s. Every version is evaluated before it ships.

Kaba remote compute
Run.Observe.Train.Repeat.

A loop that runs continuously, stays on your hardware, and compounds with every task.

Everyday

Intelligence off.
Everything else on.

Models and memory are one part of Kaba, and every bit of it can be switched off. What stays is the best way you have ever used the internet: private, peer to peer, and built with features the big browsers will never ship.

Kaba browser
A private browser
Your browsing stays on your device. No ad model, no telemetry, nothing sold.
Ghost Mode
Ghost Mode and Tor, per pane
Open any pane over Tor or in Ghost Mode, right beside your everyday tabs.
A mesh VPN, built in
Every device you own joins one encrypted private network by default. No accounts with a third party, no cloud broker, no open ports.
Route through your peers
Send traffic through any peer as an exit node: your home connection from a café, or an office server from the road.
A terminal multiplexer
Splits, tabs, and sessions that persist, with a real shell on any peer in the mesh. Detach on one device, pick up on another.
Your devices in your pocket
Reach every machine you own from your phone, wherever you are.
Make it yours
Configuration files in the spirit of vim and i3, with an SDK on the way.
Control

You say yes. Or nothing happens.

Anything sensitive waits for your approval. Approve it and Kaba learns from your choice. Deny it and nothing leaves your machine.

Happy Kaba hippo
Approved
Sent, in policy, on the record.
Your choice trains the next version of your expert.
Angry Kaba hippo
Denied
Nothing left the sandbox.
The network stayed closed and your data stayed home.

Which Kaba is for you?

Your hardware.
Your model.

$ curl -fsSL https://kaba.ai/install.sh | sh
Kaba Personal

Be your own lab.

Your laptop, desktop, and home lab become one private intelligence that learns from how you work. Build datasets from your own context, train LoRA experts on your own hardware, and own every one of them. Local and encrypted by default.

your/research-notes.lora· owned by you your/code-review.lora· owned by you your/writing-voice.lora· owned by you
On this page {{ t.label }}

The client is the lab

The client docs ↗

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.
LoRA yours browser terminal files harness
Navigation
Navigation
Every peer, model, memory, and run in one workspace. Jump back into frames where you left off.
Model harness
Model harness and lab
Chat, tool loops, and adapters over any local model. Every exchange becomes trajectory data that sharpens the next LoRA.
{{ tile.title }}
{{ tile.placeholder }}
{{ tile.tag }}
{{ tile.title }}
{{ tile.desc }}

How Kaba works

Protocols docs ↗
// THE MESH — peers, streams, tokens
laptop laptop.kaba gpu-tower gpu-tower.kaba k8s-node homelab.kaba phone phone.kaba tor egress kaba/infer/v1 kaba/train/v1 kaba/fs/v1 kaba/sync/v1 kaba/pty/v1
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.
kabactl — zsh
$ kabactl cluster join --ticket kaba-intro
✓ gpu-tower joined · kaba/cluster/v1
$ kabactl infer --peer gpu-tower
▸ streaming tokens · kaba/infer/v1
$ kabactl train lora --dataset memories/
▸ checkpoint 3/12 · trajectory logged
▸ policy: local-only · encrypted at rest
// THE LEARNING LOOP — observe → memory → RL → adapter
observe attention memory memcrypt RL trajectory LoRA adapter yours, versioned
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 parallel tasks → peer GPUs 05 hooks gate every edit

Marketplace — an open market for experts

Marketplace docs ↗
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.
rust-dev .lora · yours sec-audit .lora · leased offspring safe-rust.lora bred, versioned, yours to list
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.
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.
Tools for the loop
Ship the tools you design for your experts — sandboxed, gVisor-contained, ready to drop into anyone's tool loop.
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.

Mobile — your cluster in your pocket

Mobile docs ↗
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.

kabactl — the whole mesh from one prompt

Kabactl docs ↗
kabactl
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
$
Composable
Reads stdin, writes JSON. Every command pipes into the next. Your mesh speaks Unix.
Peer-addressed
Target any device you own by name. LAN, tailnet, or across the world — the stream finds it and connects you directly.
Same rules apply
The CLI goes through the identical policy gate as the client. No backdoor, no service account, no exceptions.

How Kaba compares

Open-source harnesses and local AI tools, compared on what Kaba adds: memory from your own activity, training on it, policy on every step, and a private mesh across your devices.

Feature
Kaba Personal
Harness and lab
{{ c.name }}
{{ c.kind }}
{{ r.label }}
{{ v.note }}
{{ v.note }}
Compare Kaba Personal with
Kaba Personal
{{ cmpPOther }}
{{ r.label }}
{{ r.k.text }}
{{ r.o.text }}
Built in Partial or via add-ons Not offered or not documented Based on each project's public documentation, October 2026. Product names are trademarks of their owners.
{{ qa.q }}
{{ qa.a }}

Your hardware. Your model.
Start training your attention.

KABA PERSONAL
Install in one line.
$ curl -fsSL https://kaba.ai/install.sh | sh
Download for {{ osLabel }} free · open source
Kaba Enterprise

Today, prompts, files, and the judgment of your best people leave through tools nobody approved.

Intelligence your company controls,
end to end.

Put intelligence to work on your own infrastructure, with your people, ontology, policies, and expertise in every loop. Accelerate adoption without turning speed into insider risk.

Book a 30-minute briefing Read the docs
A team from GE CIRT and Mandiant R&D, led by the founder of Threat Stack, Critical Stack and Kolide.
On this page {{ t.label }}
Why a harness

Every team is already putting intelligence to work. The question is whether the company can see it, govern it, and keep what it learns.

Without a shared foundation, each new tool brings its own data copies, its own connectors, its own bill, and its own blind spots. Adding the next one never gets easier. A harness gives every model, workflow, and team the same ground to stand on, so the tenth deployment costs less than the first.

Security

Data, prompts, and weights stay on systems you own. Kaba acts under the same identity, access rules, and keys as the person it works for.

Regulation

Know where data went, which model touched it, and why. Audit trails, residency, and retention controls support HIPAA, GDPR, and SOC 2 programs, ready when the auditor asks.

Risk

Automation moves fast. Policy decides what they can do, people approve what matters, and nothing ships to production on speed alone.

Cost

Every run is metered by team, project, and model. Budgets are enforced by policy, so spend is planned instead of discovered on the invoice.

Company knowledge

Build your knowledge base and models once, and let every new team and tool inherit them. Expertise compounds instead of being rebuilt each time.

Fleet management

Enroll every device, push policy and updates centrally, and revoke keys and data remotely. One console for every machine running Kaba.

Fleet search

Work done across the fleet is indexed with its permissions intact, so people find what colleagues already solved instead of solving it again.

Model pipelines

Turnkey pipelines take approved company data to trained, evaluated expert models, with clean training sets and no contamination between teams.

See it on your infrastructure, with your policies and approvers in place from day one.

A team from GE CIRT and Mandiant R&D, led by the founder of Threat Stack, Critical Stack and Kolide.
Design partners
Partner logo
Partner logo
Partner logo
Partner logo
Customer quote goes here, ideally from a CISO or head of IT, about control or cost.
Name, title, company
Public sector

Super Intelligence (SI), on hardware you control.

Government requirements shaped Kaba from the start. Data sovereignty, classification boundaries, and full auditability were design inputs, so agencies can run Super Intelligence (SI) under their own authority.

Kaba runs 100% air-gapped. Models, data, and training stay on government hardware, with no outside connection required, and every run is governed by policy, audited, and accounted for.

Kaba is written in Rust, a memory-safe language, and follows the memory safety guidance published by NSA and CISA for software used across the Department of Defense and federal government.

An operating system for intelligence.

Central control. Distributed compute.

An operating system decides what runs, what it can access, and how it talks to the outside world. Kaba does the same for intelligence across every device in the fleet, through the harness and its private network.

Intelligence makes thousands of small decisions, and every round trip to a remote server slows each one down. So nothing is virtualized and no screen is streamed: work runs natively on each device, or the nearest peer in your mesh, while IT keeps enrollment, central policy, access rules, updates, audit, and remote wipe of keys and data.

Processes
Every task and tool call, sandboxed in gVisor
Permissions
Policies on what can infer, train, read, and leave, traveling with every workload
Users and roles
People, approvers, and ontology rules that steer every action
File system
Encrypted memories, datasets, and adapters that stay where they were created, with provenance
Networking
Private iroh mesh with key-derived .kaba names, no open ports
Package manager
Models, adapters, and tools, versioned and pushed per fleet group
Fleet
Devices enrolled and grouped, updated and wiped remotely, audited centrally

Enterprise in the loop

Enterprise docs ↗

Every deployment runs through three loops: your people and ontology, your policies, and your expertise. They are built into the runtime, not sold as add-ons.

a)
People and ontology steer
Kaba acts inside the structure of your business. Steer it with people, with your ontology, or both: approvers sign off on sensitive actions, and the ontology defines the entities, relationships, and rules every step must respect. Each step is checkpointed as a trajectory, so any run can be replayed and audited.
b)
Policy inside every call
Policies run in the daemon, not in a dashboard: who can infer, what can train, where data rests, which tools may touch the network. Nothing leaves unless a policy allows it, and unsanctioned models have nowhere to run.
c)
Expertise stays in the company
The judgment of your people is captured as LoRA adapters trained on their own work. Adapters are versioned, owned by you, and never shared with a model vendor.
approvals — legal-ops.kaba
awaiting sign-off
Kaba wants to
Send a redline summary to counsel at a partner firm
policylegal.external-share · 1 approver
ontologyContract › Counterparty · rule passed
data3 docs · confidential
expertlegal-redline.lora · v14
sandboxgvisor · network held
trajectory #48213 · 6 steps checkpointed
Happy Kaba hippo
approved · sent
Redline summary sent to counsel at the partner firm
Approved by you, in policy, on the record.
approveryou · just now
policylegal.external-share · satisfied
sent3 docs · encrypted in transit
networkopened for this step, then closed
learningapproval saved for legal-redline.lora v15
trajectory #48213 · 7 steps checkpointed · complete
Angry Kaba hippo
denied · blocked
Nothing left the sandbox
The request was stopped before a single byte went out.
approveryou · denied just now
policylegal.external-share · blocked
data3 docs · stayed on device
networkheld · never opened
learninglegal-redline.lora will ask first next time
trajectory #48213 · 7 steps checkpointed · stopped at step 7

Run the whole fleet from one place

Every laptop, workstation, server, and edge device running Kaba is part of one fleet. Manage it, search it, and train across it, with every result feeding the company model.

Fleet management
Enroll devices with join tickets and group them by team, site, or business unit. Push policies, models, and adapters to each group, and see health, versions, and GPU capacity for every peer in one view.
Fleet search
Search memories, documents, trajectories, and adapters across every device without copying data into a central index. Queries run where the data lives and return only what policy lets the requester see.
Fleet training
Schedule jobs across idle GPUs in the fleet. Go beyond LoRA: distill larger models into smaller ones, run RL on trajectories and approvals, tune the router, and evaluate every candidate before it ships.
One pipeline from fleet activity to the company's mixture-of-experts model.
Collect
Trajectories and approvals from every device
Curate
Policy-scrubbed, provenance attached
Train
LoRA, distillation, RL, router tuning
Evaluate
Evals and sign-off gate every release
Roll out
New experts to fleet groups
↺ Back into the company model, and around again

A look inside Fleet Control

Sign-in through your SSO and directory, devices and the policies each group receives, read-only views of live sessions, a review queue for held files and prompts, and what the fleet costs to run.

INTERACTIVE PREVIEW · SAMPLE DATA
Kaba Fleet Control
Fleet Control JTF Cyber East
kaba://fleet/?view={{ fc.view }}
▲ 1,284 online ⚑ {{ fc.pendingCount }} pending GPU 61%
MR
Thursday, October 8 · 09:42 EDT

Fleet overview

{{ k.label }} {{ k.value }} {{ k.note }}
Needs attention Last 24 hours
{{ a.title }}
{{ a.meta }}
{{ a.time }}
Sites
{{ s.name }}{{ s.online }}
{{ s.meta }}

Identity

Who can sign in to Kaba, and which groups they land in. Group membership decides the policy a user's devices receive.

LDAP directory ● Bound · synced 6 min ago
Group mapping Directory group → Kaba group
{{ g.dn }} → {{ g.group }} {{ g.users }}
Users in no mapped group are denied sign-in. CAC/PIV is required for all groups.

Devices 1,312

Search hostname, user, serial…
HostnameGroupHardwarePolicyStatusSeen
{{ d.host }}{{ d.user }}
{{ d.group }} {{ d.hw }} {{ d.policy }} {{ d.status }} {{ d.seen }}

Policies & assignment

Toggle a cell to assign a policy to a group. Changes are staged until you push.

{{ fc.stagedCount }} staged
Policy
{{ g.name }}{{ g.count }}
{{ row.name }} {{ row.desc }}
{{ pd.k }}
{{ it.k }}{{ it.v }}
Showing rules for {{ fc.selectedPolicyName }}. Staged rollout: 5% canary for 30 min, then the rest of the group.

Live sessions 214

Read-only view of inference sessions. Every observation is logged to the user's audit trail.

{{ c }}
{{ fc.sess.user }} · {{ fc.sess.title }} {{ fc.sess.meta }}
{{ fc.sess.stateLabel }}
{{ t.who }}
{{ t.text }}

Review queue

Files and prompts held by policy before reaching the model or leaving the device.

{{ r.sev }}
{{ r.file }} {{ r.rule }} {{ r.user }} · {{ r.time }}
{{ r.status }}
{{ fc.rev.file }} {{ fc.rev.user }} · {{ fc.rev.device }} · {{ fc.rev.direction }}
{{ fc.rev.status }}
{{ h }}
{{ ln.n }}{{ ln.pre }}{{ ln.hit }}{{ ln.post }}
Chain of custody
{{ lg.t }}{{ lg.e }}

Cost & performance

{{ k.label }} {{ k.value }} {{ k.note }}
Tokens per day
On-devicePeer
{{ fc.rangeStart }}Today
Spend by group
{{ sg.name }}{{ sg.amt }}
ModelDevicesRequestsp50 latencyp95Tokens/s
{{ m.name }} {{ m.devices }} {{ m.req }} {{ m.p50 }} {{ m.p95 }} {{ m.tps }}
≡ KABA FLEET CONTROL · {{ fc.viewUpper }} POLICY v{{ fc.policyVersion }} · AUDIT LOG ON · FIPS 140-3
Names, sites and figures are illustrative.

A model that builds itself from your people

Each team trains adapters on its own work. Kaba routes across them as a mixture of experts, so the company model improves every time someone does their job well.

Legallegal-redline.loraEngineeringinfra-runbook.loraSalesdeal-desk.loraSupporttier2-triage.loraFinanceclose-checklist.lorarouterMoEcompanyexpertpeopleadapters they trainacme.kaba
Routing learns from outcomes
Trajectories and approvals tell the router which expert answers best, per task and per team.
Every expert is accountable
Adapters ship with a manifest and eval trail: who trained it, on what, and how it scored.
Retire, retrain, or merge
Swap experts in and out without retraining the base model. Merge two adapters into a new one.

Speed without understanding is an insider threat

Security docs ↗

The fastest way to adopt intelligence is to let everyone use it. When nobody understands where data goes, what automation can touch, or which models learn from it, that speed becomes the biggest risk inside your company. Kaba lets you accelerate with the risk visible and governed from the first deployment.

Sanctioned by default
Every model call and tool step routes through policy. RAG, models, adapters, and memories are encrypted at rest. Unapproved endpoints are blocked at the daemon.
Every tool sandboxed
Each exec runs in a gVisor container. Network access is a policy decision, granted per step.
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.
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.
Deploys on your platform
Kubernetes, Nomad, or any container cluster you run. Carve the mesh per user, per business unit, per org: same binary, scaled on your systems.
Verifiable by construction
Open source binary you can inspect, plus a trajectory for every run. GDPR and CCPA answers you can prove. Runs fully air-gapped when required.

You can't control the cost of intelligence until you control its use

For most companies, spend on intelligence is growing faster than anyone can track. Seats, API tokens, and cloud GPUs are billed by different vendors to different teams.

Cost and risk are the same problem. When you can't see who is using which model on what data, you can't secure it and you can't budget for it. Kaba puts use under control first, and cost control follows.

Per seat and per token
The bill grows with every hire and every prompt.
Seats, API tokens, and cloud GPUs from several vendors, each billed to a different team.
Kaba on hardware you own
Usage grows. The bill stays where you planned it.
Small models with your adapters on machines you already run, plus a shared GPU pool on one meter.
Small models on your systems
Most work doesn't need a frontier model. Small models with your own adapters run on hardware you already own, with no per-token bill, and every request that stays in-house is one less exposure.
Dedicated GPU boxes, one meter
Put GPU capacity in a few dedicated machines and share them across the fleet over the mesh. Spend and utilization are monitored at a single point instead of spread across vendors and teams.
Budgets enforced as policy
Set who can use which models, how much compute each team gets, and when to fall back to a smaller model. Limits are enforced in the runtime, not reconciled from invoices at the end of the month.
Every run accounted for
Each run is recorded with the model, the hardware it ran on, and the team it served. Cost is attributed from the same trajectories you use for audit.

How Kaba Enterprise compares

Compared with platforms enterprises evaluate for governed intelligence. Prisma AIRS secures AI traffic and can run alongside Kaba; Palantir AIP grounds models in a central data platform; Microsoft 365 Copilot runs in Microsoft's cloud.

Feature
Kaba Enterprise
Operating system for AI
{{ c.name }}
{{ c.kind }}
{{ r.label }}
{{ v.note }}
{{ v.note }}
Compare Kaba Enterprise with
Kaba Enterprise
{{ cmpEOther }}
{{ r.label }}
{{ r.k.text }}
{{ r.o.text }}
Built in Partial or via add-ons Not offered or not documented Based on public product documentation, October 2026. Product names are trademarks of their owners.

See it on your infrastructure, with your policies and approvers in place from day one.

A team from GE CIRT and Mandiant R&D, led by the founder of Threat Stack, Critical Stack and Kolide.
{{ qa.q }}
{{ qa.a }}

Your hardware. Your model.
Start training your attention.

KABA ENTERPRISE
Bring Kaba to your company.
A self-hosted pilot on your infrastructure, with your policies and approvers in place from day one.
A team from GE CIRT and Mandiant R&D, led by the founder of Threat Stack, Critical Stack and Kolide.
About Kaba Labs

Security people who don't want to slow you down.

We've spent decades in enterprise security. We've also watched security teams become the reason new technology stalls.

We don't think that trade-off is necessary. You can move fast, understand your risk, and protect your data at the same time. Kaba is how we're proving it.

Our team has worked incident response at GE CIRT and done R&D at Mandiant. Our founder started Threat Stack, an early cloud and Linux security company built on deep syscall monitoring and acquired by F5, and Critical Stack, a secure Kubernetes platform acquired by Capital One, and co-founded Kolide, the endpoint monitoring company built with the team behind osquery and acquired by 1Password. Kaba is the latest of those, and not the last.

Based in
Boston, MA
Background
GE CIRT · Mandiant R&D · founder of Threat Stack (acquired by F5) and Critical Stack (acquired by Capital One), co-founder of Kolide (acquired by 1Password)
Building
The harness and lab for intelligence you own
Accessibility is a Human right. Your data. Your model. Your experience. Real-time and continuously verified.
© Kaba Labs, Inc 2026
DOWNLOAD DOCS ABOUT SOURCE PRIVACY
COMING SOON
[click anywhere to close]
Screenshot
[click anywhere to close]