Yantrik

Yantra — an instrument of control

A mind that works for you.
A memory it keeps.
A desktop it acts on.

Yantrik is a small family of open-source pieces that fit together into one instrument. YantrikDB remembers. Yantrik Mind thinks and acts. Yantrik OS is the desktop it acts on, and the door any other agent comes in through. You hold it.

Start with the memory Watch the desktop, live Nightly ISO

It is early. YantrikDB is in daily use and on PyPI and crates.io; the OS ships a nightly build; the Mind is being built in the open. Everything below links to the real thing.

The Yantrik instrument: you at the centre, minds, memory, the desktop, and the gates agents enter through

The Yantrik emblem, read inside-out. Hover a layer, or scroll.

01 · the bindu

You hold it.

Everything in Yantrik converges on one person. Minds work for you in Mind View, a view of their own: they cannot move your windows or take your focus. One known exception, being fixed: if your own browser was open first, a Mind's web tools can reach it, so don't leave it signed in while a Mind works. In the default mode, anything destructive asks you first. So do actions on the OS and sending email, each with a card showing the full text. Not every route off the machine asks yet; the rest are being put behind the same card, in changes that are open and not yet merged. What a Mind did is on record afterwards, with the tool it ran, how long it took, and what it touched.

The Mind View desk bar: Yantrik Mind, running a tool, 'Working on it', the describe tool reported and running, 0:02 elapsed, 1 call, 0 failed; the footer reads Browser, own account, network checked, OS actions ask first.
The desk bar while a Mind works: what it is doing, how long, and how many calls. A real capture.

02 · the lattice

Minds that remember what they believe.

Yantrik Mind is a Rust agent built on YantrikDB's typed memory. It does not keep notes as flat text: it keeps beliefs, each with a confidence score, an evidence trail and, where two disagree, a contradiction edge. New evidence revises confidence the Bayesian way. When it researches a question it recalls what it already believed, checks it against live sources, and tells you when the two disagree rather than asserting either. Conversations are consolidated into durable beliefs instead of being truncated.

Everything it can do sits behind one deterministic harm gate. Sub-agents can propose an outward action; none of them can confirm one.

It is not the only mind. Claude Code, Hermes, OpenClaw and Pi use YantrikDB too, through an MCP server or a Hermes plugin, and reach the same desktop through the gates.

yantrik-mind on GitHub

03 · the strata

One memory, kept on your machine.

YantrikDB is a local-first cognitive memory engine: one SQLite file on your machine, built so every mind there can share it. On today's Yantrik OS image each harness is still installed with its own store, and the shared mode is in progress. It consolidates many notes into one fact. It notices contradictions and keeps both sides until someone decides. It keeps history, so you can ask what it believed in March. And it explains why it recalled something, including when the memory is old or rarely confirmed. It is a Rust library with Python bindings, an HTTP server, an MCP server for Claude Code, Cursor and Windsurf, and a memory provider for Hermes.

Of the pieces here, it is the one with the most use today.

# Python
pip install yantrikdb
# Rust
cargo add yantrikdb

yantrikdb.com · PyPI · crates.io · version 0.23.4, stars as of 2 Nov 2026

04 · the gear

A desktop built to be driven.

Yantrik OS is an AI-native Linux desktop. It is a real desktop first: the Bench, with your files and folders on it; a launcher; apps on one framework (19 of its 21 apps today), so Notes, Calendar, Weather, Terminal and the rest behave alike; updates that are signed (ed25519), verified before install, and rolled back with one command. What makes it different is that every app publishes a control surface, so a Mind can read it and act on it the way you would, and you can watch.

It is early. Builds are nightly, and rough edges are part of the deal. You can boot the ISO in a VM, or watch a machine running it right now.

Yantrik OS: the launcher open over the Bench. The desktop greets Alex, shows Notes, Terminal, folders and files, and a dock; the launcher lists 35 apps including Agents, Calendar, Memory, Notes, Permissions and Personality. The status bar reads Mind · Asks first.
The launcher over the Bench, on the current nightly. Sample data; a real capture.
Four Yantrik OS apps: Calendar in week view with three events, Notes showing a Weekend hike note, Weather for Seattle, and Files on the Desktop folder. Each app has a Context panel with what a Mind could do next.
Calendar, Notes, Weather, Files: the same framework under most of them, and a Context panel in each that says what a Mind could do next.

yantrikos.com · live demo · nightly ISO · GitHub

05 · the gates

One door for every agent.

Every yantra stands inside four gates. Here they are the control surface: one way in, for any agent. yos describe returns what an app is showing, as text and as JSON with a revision. yos act changes it. Claude Code, Hermes, OpenClaw, Pi and Yantrik's own Mind all use the same door, so the rules at the door apply to all of them, and what each one did is on record.

Left: the Weather app showing 9 degrees in Seattle with the week's forecast. Right: a Terminal running yos describe weather, which prints the same conditions as a summary line and JSON with a revision id.
The Weather app, and yos describe weather reading it from a terminal: the same facts, a revision id, no screenshot needed.

How a task moves through it

  1. You ask.In the Mind's own view, or from a terminal, or from Claude Code on another machine.
  2. The Mind recalls.YantrikDB returns what it knows, how sure it is, and why, with its contradictions attached.
  3. It reads the desktop.yos describe: what each app shows, as data.
  4. It acts.yos act, through the same door as every other agent.
  5. Outward asks.A card with the full text, and your answer. Today for OS actions and email; the other outward tools are being gated the same way.
  6. It remembers.What happened is consolidated into memory, so the next mind starts where this one stopped.

Research

Tier: how a small model finds the right tool.

A 1.5B model shown all 80 tools picks the right one half the time. Give it the right neighbourhood and it picks the right tool 89% of the time. The bottleneck is not selection, it is navigation. Tier adapts how tools are presented to the size of the model: one tool definition, a different face per tier. On sub-4B models that is +10 points of accuracy and 97% fewer tool tokens. It matters here because a desktop with 35 apps is a lot of tools, and not every mind is large.

tier on GitHub · the paper (Zenodo) · pip install yantrikos-sdk

ModelRight tool, all 80 shownRight tool, given its family
1.5B50%89%
9B—98%
35B88%98%

Figures from the Tier README, on an 80-tool suite; the 9B all-tools figure is not reported there.

Where each piece stands

As of 2 November 2026. Honest, because you will find out anyway.

PieceWhat it isStateWhere
YantrikDBlocal-first cognitive memory0.23.4 on PyPI and crates.io; four repos, 369 starsyantrikdb.com
Yantrik OSAI-native Linux desktopearly; nightly ISO; a live machine to watchyantrikos.com
Yantrik MindRust agent on typed memoryunder construction, in the openGitHub
Tieradaptive tool routingpaper on Zenodo; SDK 0.3.0 on PyPIGitHub