SkillAgentSearch skills...

yantrik-os

AI-native Linux desktop where your AI agents (Claude Code, Hermes, OpenClaw, Pi) drive every app, remember with YantrikDB, and show proof of what they did. Boot the live ISO.

Install / Use

claude mcp add yantrikos -- npx -y github:yantrikos/yantrik-os

If the server publishes to npm under a different name, use that package instead — check the repo README.

About this skill
🔌

MCP Server

Model Context Protocol server

Quality Score

84/100

Supported Platforms

Claude Code
Claude Desktop

Our assessment of yantrik-os

yantrik-os scores 84/100 on our quality scale, 609th of 947 AI & Machine Learning skills we index.

Its MCP Server is 25 KB long, well organised into 22 sections with 11 code examples: a thorough specification that gives an agent plenty to work with.

It has 10 GitHub stars, so there is little community track record yet; judge it on its content.

Substance
30/30
Structure
20/20
Description
15/15
Adoption
4/20
Freshness
15/15

Maintenance, license and trust

  • The repository was last updated today, so yantrik-os is actively maintained.
  • It is released under GPL-3.0, a copyleft license: you can use it, but modified versions you distribute must carry the same license.
  • Its trust signals score 97/100, with no cautions. These come from repository metadata, not a code audit — read the skill file before letting an agent act on it.

Safety scan

No issues found

Our scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands. An AI review of the same text found nothing harmful.

AI review by kimi-k2.7-code on 2026-11-16. Automated pattern scan on 2026-11-16. It catches known dangerous patterns, not every risk — read a skill before letting an agent act on it.

yantrik-os compared with similar skills

All 4 of these similar skills score higher than yantrik-os; compare them before choosing.

SkillScoreStarsUpdatedFormat
yantrik-os (this skill)by yantrikos8410todayMCP Server
claude-memby thedotmack100117.6ktodayCLAUDE.md
Agent-Reachby Panniantong100112.3k8d agoCLAUDE.md
Understand-Anythingby Egonex-AI10092.8k1d agoCLAUDE.md
rufloby ruvnet10081.9ktodayCLAUDE.md

Frequently asked questions

How do I install yantrik-os?
Run claude mcp add yantrikos -- npx -y github:yantrikos/yantrik-os. The install tabs above show the steps for each supported agent.
Which AI agents does yantrik-os work with?
It is written for Claude Code and Claude Desktop, as a MCP Server file. Other agents that read the same format can often use it too.
Is yantrik-os safe to use?
Our scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands. An AI review of the same text found nothing harmful. It is GPL-3.0-licensed and scores 97/100 on trust signals. Skills are instructions an agent will follow, so read the file before installing it and do not approve commands you do not understand.
Is yantrik-os still maintained?
The repository was last updated today, so yantrik-os is actively maintained.

Yantrik OS

Chat on Discord Nightly image CI License: GPL-3.0

An AI-native desktop operating system where the AI is the shell. Built in Rust, on Debian 13, shipped as a live ISO you boot and try before you install anything.

Bring your own AI agents (Claude Code, Hermes, OpenClaw, Pi, or the built-in Yantrik Mind) and they drive the whole desktop: every app and every screen, through the same path a person uses. What they learn about you stays in YantrikDB on your machine, and the desktop keeps a record of what each one did, so you never have to take an agent's word for it.

Come and talk to us: discord.gg/7cDw3jd3Xf — #nightly announces every published image, #apps-and-surfaces is where the app control protocol gets argued about, and the help forum is the place to say what broke. It is a small server and a question there is read by the people who wrote the code.

What makes it different

  • The AI is the shell, not an add-on. The companion is woven through file management, email triage, presentations and system monitoring rather than sitting in a chat window beside them.
  • Every screen is drivable by an agent. The shell and every app publish their state and accept actions over a unix socket — yos describe shell, yos act shell open_app name=notes. Anything a person can do from the keyboard, an agent can do through the same path. See docs/app-control.md.
  • Any mind, and the desktop says which one. A local model through Ollama or llama.cpp, a cloud provider, or a separate agent that attaches — all equally first-class. The OS itself sends nothing anywhere: no telemetry, no phone-home, no call it makes on its own. But it is built to be driven by whatever mind you give it, and a mind you point at a provider talks to that provider. The status bar carries a chip naming which mind is answering and whether it is local or cloud, so that is never something you have to go and check.
  • Proof of what your agents did. Each mind works under a mode you choose: read-only, asking about anything that matters, or asking only before what cannot be undone. Whatever it does without asking is written down where you can read it (the mode chip's "See what it did without asking").
  • It speaks up only when it is worth it. A four-stage pipeline (Detect → Generate → Score → Deliver) decides when a suggestion is worth your attention, so it helps without nagging.
  • Rust throughout. Slint UI, agent, memory database and system observer. No Electron, no Python runtime, no Docker.

Built-in apps

Sixteen application binaries ship, each its own window, each drivable by the agent. Two more exist in the tree and are shelved — deliberately not in this build, and listed below so that this page cannot promise them.

| App | What it is | |-----|-----------| | Notes | Markdown notes with semantic search, backlinks, versions | | Email | IMAP client with AI triage and smart notifications | | Calendar | Events, week and day views, reminders | | Weather | Current conditions, hourly and daily forecast, alerts | | Terminal | Terminal emulator with AI command assist and split panes | | Editor | Text editor with tabs, find and replace, go-to-line | | yDoc | Document editor — rich text, comments, change tracking | | yPresent | Presentations — AI deck generation, templates, speaker notes | | Images | Viewer with zoom, rotate, crop, slideshow, batch operations | | Downloads | Resumable transfers with checksum verification | | Snippets | Code snippets by language and collection | | Containers | Docker/Podman containers, images and volumes | | Network | WiFi, ethernet and bluetooth | | System Monitor | CPU, memory, disk, network, processes | | Arcade | Builds a playable game from two small JSON specs, verified headlessly | | Studio | Pictures from a sentence — your own GPU through ComfyUI, a hosted service, or neither |

Blender, if it is installed, is drivable too — and it is not ours. The launcher starts it with an addon that binds the same kind of socket every app here does, so yos describe blender reports the scene, the objects, the engine and the resolution, and yos act blender add_primitive, set_light, set_camera and render work from a terminal or a mind exactly as an app's own actions do. render is sensitive and run_python is dangerous, for the obvious reasons. This is the first thing on the desktop that shows the control protocol is not a house convention: a program nobody here wrote can join it by publishing one socket.

Shelved — not in this build:

| App | Why | What brings it back | |-----|-----|---------------------| | ySheets | No cell model behind the grid, so nothing could be typed into it, by mouse or by mind. No formula engine at all | A cell model, CSV load and save, and arithmetic with references plus SUM/AVG/MIN/MAX/COUNT | | Music | Nothing plays audio — no playback engine, no scanner and no library behind the screen | mpv driven over its JSON IPC socket, a folder scan filling a library store, and play/pause/next/queue and open <file> working |

Both crates stay in the tree and keep compiling; what changed is that a person cannot reach them and a mind is not told they exist. Asking for one gets a refusal that says why, not "unknown app". The account is design/shelved-2026-09-20.md and the list every surface reads is SHELVED in crates/yantrik-ui/src/wire/dock.rs. (Audio does play — the shell's own media screen drives a real mpv. That was never the Music app.)

The shell itself provides Files, Settings, Memories, Notifications, Bond, Personality, Permissions, Devices, Packages, Skills and About as screens rather than separate windows.

Every one of them wears the same frame. See docs/app-sdk.md for why that is structural rather than a convention anyone has to remember.

The companion

Not a chatbot. A proactive agent with:

  • Instincts — email watch, open loops, routine learning, commitment tracking, security
  • Bond — a relationship that moves from Stranger through Acquaintance, Companion and Confidant to Partner, based on the quality of the interaction
  • Memory — persistent vector-indexed recall that grows over time
  • Model-adaptive behaviour — detects what the model can do and adjusts tool use and prompt complexity to match
  • Pluggable minds — the built-in companion is one harness among several. Anything that speaks the attach protocol can answer instead, managing its own endpoint and credentials. Five exist: Yantrik Mind, Hermes Agent, Pi, DeepSeek and OpenClaw — the last four ship as source in harnesses/, and none of them is started until you configure it. See docs/harness.md.
  • YAML plugins — add tools without writing Rust

Architecture

┌────────────────────────────────────────────────────────────┐
│                        Yantrik OS                          │
│                                                            │
│   yantrik-ui ──────── yantrik-companion ──── yantrik-ml    │
│    (shell)              (agent)               (inference)  │
│        │                    │                              │
│   yantrik-os          yantrikdb                            │
│    (system)            (memory)                            │
│                                                            │
│   16 app binaries · 10 services · one control surface      │
│                                                            │
│   Debian 13 → labwc (Wayland) → Slint                      │
└────────────────────────────────────────────────────────────┘

25 crates, 16 shipped apps (18 in the tree, two shelved) and 10 services. The ones worth knowing:

| Crate | Purpose | |-------|---------| | yantrik-ui | The shell — Slint UI, app wiring, the control surface | | yantrik-companion | The agent — tools, instincts, bond, personality, proactive pipeline | | yantrik-ml | Inference — LLM backends (Ollama, OpenAI-compatible, llama.cpp, Claude CLI), embeddings, STT/TTS | | yantrik-os | System integration — D-Bus, inotify, sysinfo, battery, network, processes | | yantrik-harness | The attach protocol a third-party mind implements to answer for the shell | | yantrik-ui-kit | The UI kit every app draws from, including the mandatory AppHeader | | yantrik-app-runtime | What an app binary is built on — instance guard, theme, IPC, control surface | | yantrik-design-tokens | Colour, type, spacing and size tokens, shared by the shell and every app |

Threads: the Slint event loop, a system observer (D-Bus, file watches, polling) and a companion worker (inference, memory, tools).

Quick start

Install

Yantrik OS is a Debian 13 (trixie) live ISO, built and boot-tested by CI. Download it, check the checksum, boot it; the disk installer is inside the running image.

Images are at https://iso.yantrikos.com/nightly/, where latest.json names the current file, its sha256, its size and its version.

sh install.sh               # prints url, size and sha256 of the current image, and stops
sh install.sh --download    # also fetches it here and verifies the sha256

With no flag it installs nothing, writes nothing and wants no privilege — including when it is piped into a shell, which is the way it is most often run.

Only the nightly channel has ever had a build published to it. beta and stable are names the updater knows and nothing has been published to either.

Once it is booted, yantrik-install in the live session writes it to a disk. docs/getting-started.md is the whole path.

There is also deploy/yantrik-os/cloud-init/user-data.yaml for Proxmox, libvirt or anything that takes a cloud-init file. That is how the project's own test machines are made, not how a person installs this.

Hardware

The image is about 1.31 GiB. The Minimum column is what the first-boot hardware scan checks — the disk figure against the whole size of the disk the installer will write to, not against free space on the live session. The other two columns are what the image has actually been booted on:

| | Minimum | CI's boot test, every published image | The project's test machine | |--|---------|---------------------------------------|----------------------------| | CPU | 2 cores | 4 | 4 | | RAM | 4 GB | 4 GB | 8 GB | | GPU | none | none (virtio-vga, software rendering) | none (virtio-gpu, software rendering) | | Disk | 6 GB, to install | none — boots live | 32 GB | | Firmware | BIOS (UEFI is built, untested) | BIOS | BIOS |

Real hardware, the UEFI path and any GPU other than QEMU's are not measured — nobody has checked. docs/hardware-requirements.md says exactly which numbers are measured and which are not.

Without a GPU the desktop is fully usable and inference is slow. That trade is deliberate: the shell renders through Slint's software rasteriser and idles at around 2% of a core.

LLM backends

The image ships no model and no mind. You point it at one during first-run setup.

| Backend | Setup | Where what you type goes | |---------|-------|--------------------------| | Ollama | Point at a local or remote Ollama server | That machine. Nothing leaves your network

Truncated for display — read the full file on GitHub.

Related Skills

View on GitHub
GitHub Stars10
CategoryAI
Updated7h ago
Forks8

Languages

Rust

Trust signals

97/100

From repository metadata: license, adoption, age and documentation. Not a code audit — see the Safety scan above for what the skill file itself contains.

1 info