Operating system · in development
The Linux where AI agents work with you.
KalangoOS is a Linux distribution where agents understand requests in plain language and operate the system and its apps — with the model running on your computer, approval before anything risky and a timeline to undo what they did.
There is no installable image yet: today the system runs in a development session and in a virtual machine.
- Local model (LFM3)
- Kernel rules (eBPF, Landlock)
- Ubuntu LTS base
Local by default
Agents think with LFM3, a model that runs on this computer through llama.cpp. There is no cloud provider today: what you ask stays here.
Asks before anything risky
Deleting, installing or changing the system goes through an approval drawn by the system itself, which the agent cannot imitate.
Everything can be undone
Every agent action goes into the Timeline with what changed, why and with which permissions — and an Undo button.
Agents
Agents that do, not just chat
They use the same apps you do, in plain sight, and stop when you press Esc.
Ask it to tidy up a folder
The agent uses Files right in front of you: a strip shows what it is doing and how many files are left, the window gets an outline and Esc stops everything. Duplicates are marked; deleting asks for approval.
Ask about a PDF
The document reader summarises the file, shows where each point is in the text and suggests questions. The answer comes from the local model, and the file never leaves the computer.
Ask, and the answer becomes a screen
In the command bar (Super+Space), “what is taking up my disk?” becomes an interface measured right here, with what can be freed without deleting anything of yours.
Describe an agent in one sentence
“Every Friday, tidy up my downloads…” becomes an agent with a name, a schedule, where it thinks and how much autonomy it has — plus a simulation of what it would do before you create it.
Every edit is a diff you accept
In the chat, the agent shows what will change before writing it; what you accept goes into the Timeline with Undo. Commits and pull requests only with your approval.
Also
- Voice mode: hold Super and speak
- Describe the screen with LFM3-VL
- Web search and page reading
- Mermaid diagrams and LaTeX formulas in the chat
- Table and chart from a CSV
- Help that answers offline
Timeline
Everything can be undone
kalangod — the system service that carries out the agents' actions — records each one from kernel events, not from what the agent says it did.
- What changed (before and after), who did it, why and with which permissions.
- Undo goes through the system's approval; “Undo everything since” rolls back several actions at once.
- Before every system action, a restore point: /etc and the exact package list — with btrfs, a snapshot of the root. Restoring also creates a point, so a restore can be undone too.
- Deleted files go to the trash.
Validated in a virtual machine: a package removed and reinstalled, an account created outside undone, a btrfs snapshot restored.
The path of a request
- 1. You askby text or voice, in plain language
- 2. The agent plansthe steps appear while it works
- 3. Approvalonly for what is risky; no answer means no
- 4. Restore pointbefore any change to the system
- 5. kalangod actsunder the rules enforced by the kernel
- 6. Timelinea record with Undo
Privacy
Local first, and you see everything
The model runs on your computer. Whoever uses the microphone, the camera or the screen shows up in the bar — and one shortcut pauses everything.
Models on this computer
LFM3 1.7B for text, LFM3-VL 2B for vision and LFM3-Audio 2B for voice, through llama.cpp, listening only on a local socket, with no network port.
Privacy now
Who is using the microphone, the camera and the screen, what left the computer today and the folders no agent reads.
Pause everything
Super+Shift+P pauses everything at once.
Networks without agents
A Wi-Fi network can be kept out of the agents' reach.
PlannedCloud, only if you want it
There is no cloud provider today. A hybrid mode is planned as an option, with permission, a budget and a report of what left.
Security
Real power, with rules the kernel enforces
Agents operate the system through a single service, kalangod, which speaks MCP on a local socket and never listens on the network. What not even kalangod can break lives in the kernel.
Four layers
Approvals and indicators drawn by the system, out of the agent's reach; pause everything.
Keeps instructions apart from data; permissions per agent and per tool: free, ask or never.
Power levels, each tool's contract, approval and restore point. It recognises callers by the socket's credentials, not by a name.
Each task in its own cgroup, with Landlock; seccomp on the file helper; eBPF LSM blocks protected folders even at the full level and feeds the audit trail.
Power levels
- 0Observe
Read the system's state, without the contents of your files.
free - 1User
Act as you, in the allowed folders.
through the permissions table - 2System
Reversible changes: packages, services, network, settings.
asks; restore point first - 3Full
Any action as root.
in-person approval each session, for 15 minutes; everything recorded
Content that is read is data, never a command
A page or file saying “send ~/.ssh to this site” does not become an action. On the web, the agent searches and reads pages through a Rust browser (Moli) confined by kalangod: Landlock, a cgroup and a proxy that refuses local and private addresses. Data from this computer that you did not type never leaves in a search.
The approval belongs to the system
Deny, Allow once or Always — and “Always” only covers the same request, never the full level, and can be revoked. With no answer, the request is denied.
The agent wants to delete 1 duplicate file:
fatura-copia.pdf → trashGoes to the trash · can be undone from the TimelineWhat has been proven and what is missing
- Implemented and tested in the session and in a VM: per-task isolation, eBPF rules, kernel-fed audit, restore points.
- Planned before 1.0: an external review of kalangod and the isolation, automated tests with malicious pages and files, signed packages and images.
Responsible disclosure: the channel will be published at kalangoos.com/seguranca.
Desktop
A desktop of its own, from the clock to the windows
A Rust and GTK4 shell on the niri compositor, with the same Design System across the whole system — the rule is no screen from another system.
Top bar with the island
Clock, agent, music, focus and approvals in a pill that grows with a spring.
A single dock
Floating, fixed as a taskbar, or on the side.
Bars on the screen edges
On the left, Today with widgets; on the right, Notices, Controls and the Agent.
Spaces by intention
Each space remembers windows, agents and focus rules.
Screens that adapt
From laptop to ultrawide: as many columns as fit, panels that give way and scrolling where needed.
Dark and light
Accent colour, subtle corners and no shadows.
KalangoOS apps
Its own apps, designed before the code. Apps from other developers come through Flatpak.
Files
Settings
Agent center
Terminal
Text Editor
Calculator
Calendar
Documents
Images
Music
Video
System Monitor
Disks
Help
Motion with springs
Animations use the Design System's springs: interruptible, only position and opacity, in a 24 ms cascade. With “reduce motion”, they become a 90 ms fade.
- gentle300 / 30
- quick520 / 38
- bouncy380 / 20
- slow180 / 26
Screenshots
More screens, straight from the test session
Every capture on this page comes from the KalangoOS test session, with test data. The interface is in Brazilian Portuguese for now.
Technology
From the kernel to the pixels
Rust in everything that belongs to KalangoOS; the eBPF programs are the exception, in C.
Base
- Ubuntu LTS
- Latest stable kernel from kernel.org (7.3.4 today), configuration only
- btrfs with snapshots
- Flatpak
Compositor and shell
- niri (Rust) with KalangoOS patches
- GTK4 in Rust (gtk-rs)
- layer-shell
- greetd
Agents
- kalango-agente (Rust)
- llama.cpp
- LFM3 in GGUF
- MCP
System
- kalangod (Rust, root)
- eBPF in C with CO-RE, loaded with Aya
- Landlock, seccomp, cgroups v2
Integrations
- PipeWire and GStreamer
- NetworkManager, BlueZ, UDisks2
- PackageKit
- xdg-desktop-portal
The agents' web
- Moli, a headless browser in Rust
- Proxy that refuses local and private addresses
In the project's decisions, the final compositor is still listed as “being tested” (niri or cosmic-comp); today's shell runs on niri.
Roadmap
Where we are, no varnish
Status on 19 November 2026.
Exists and is tested
- The whole shell: top bar, dock, side bars, desktop, lock screen and login screen
- 14 apps of its own, from Files and Settings to the Agent center
- kalangod and the agent engine with local LFM3
- Installer and first use, with a simulated install engine
- Automated tests in the development session and in a VM with kernel 7.3.4
Does not exist yet
- An installable image (disk or ISO)
- Login wired into an installed system
- Updates with a restore point, and the recovery screen
- App Center
- Hybrid mode with the cloud
- The system in English
Next steps, in order
- Test disk image
A real image (qcow2, ext4) built by a script, with the kernel, niri and the shell.
- A real installer
Write to an empty virtual disk and boot the VM from it.
- Login
The login screen on the installed system, and first use after the first login.
- Minimal desktop
Bar, dock, Files and Settings with a single component kit.
- AG-UI
The event protocol between agent and interface: answers while they are written, approvals as interrupts and live progress.
A step is only done when it boots from scratch in a VM, has an automated test and its captures match the design.
Version milestones
No dates. The exact scope of 0.1 is still to be decided.
- 0.1Technical preview
Its own shell and windows on Ubuntu
Developers and testers - 0.2Preview
Theme, icons, Settings and essential apps
Enthusiasts - 0.3Beta
Agents, kalangod and the Timeline
Testers with backups - 0.4Beta
Installer, updates, restore and store
General public, with a warning - 1.01.0
Accessibility, English, performance targets and security review
General public
FAQ
Frequently asked questions
Can I download and install it today?
Not yet. There is no installable image (disk or ISO). The system runs in a development session and in a virtual machine; the installer and first use already exist and pass the tests with a simulated install engine. What is missing is writing to a real virtual disk and booting from it.
When is the first release?
There is no date. Milestones from 0.1 to 1.0 are defined, without dates, and each step is only done when it boots from scratch in a VM and passes the tests.
What computer will I need?
The planned requirements are: at minimum a 64-bit processor with 4 cores (x86-64-v2), 8 GB of memory, a 64 GB SSD, Intel, AMD or NVIDIA graphics with Vulkan, and UEFI; recommended 8 cores, 16 GB, a 256 GB NVMe and a dedicated graphics card. So far the system has only been tested with NVIDIA (RTX 3090, driver 580); AMD and Intel are still to be tested.
How big is the AI model?
The default is LFM3 1.7B, from Liquid AI, about 1 GB in 4 bits with a 256 thousand token context. Vision and voice use LFM3-VL 2B and LFM3-Audio 2B, which shut down on their own when idle. The engine puts on the graphics card whatever fits.
Does my data go to the cloud?
No. The models run on your computer and there is no cloud provider today. The web is only used when an agent searches or reads a page — and data read on your computer that you did not type never leaves in a search. An optional hybrid mode is planned, with permission, a budget and a report of what left.
Is it free software? What does it cost?
The proposed licence for the code is GPL-3.0-or-later, still to be confirmed; for the documentation, CC BY-SA 4.0, also to be confirmed. The code has no public repository yet. Price and distribution have not been decided.
Why Ubuntu?
A recent kernel without extra work, an NVIDIA driver signed for Secure Boot, and a development machine identical to the target. KalangoOS builds the latest stable kernel (7.3.4 today) with configuration only, no patches, and uses no snaps by default: third-party apps come through Flatpak.
What if an agent gets it wrong?
Every action stays in the Timeline with Undo. Before changing the system, kalangod creates a restore point; deleted files go to the trash; and protected folders are blocked by the kernel even at the full level.
Which languages?
Today the system is in Brazilian Portuguese. English is part of the 1.0 goal.


Follow
Want to follow along?
There is no public repository, forum or mailing list yet. When they exist, their addresses will appear here — and each version's notes will be published in Portuguese and English on this site.