Autonomous Agent OS · Alpha RC2

The OS that thinks. On its own.

YantraOS replaces passive computing with an agent that senses its environment, reasons with LLMs, and acts continuously, in a sandboxed execution chain, on your hardware.

v1.0
Alpha RC2
6
Phases complete
MIT
Open source

An inversion of computing

Every traditional OS waits to be used. YantraOS runs a perpetual cognitive loop. You are no longer the operator. You are the principal authority in a hierarchy of agents.

Old model
HumanInputOSOutput
YantraOS
OS sensesOS reasonsOS actsHuman observes

The Cognitive Engine

The Kriya Loop

From the Sanskrit Kriya: purposeful action guided by conscious awareness. This loop runs inside yantra.service and never pauses.

01
SENSE

Sense

CPU, RAM, GPU, temperatures, logs, network. Every telemetry signal read at each cycle.

02
REMEMBER

Remember

ChromaDB vector search retrieves past context and outcomes relevant to the current state.

03
REASON

Reason

LiteLLM routes the inference call to the optimal backend: local Ollama or cloud API.

04
ACT

Act

Commands execute inside a Docker Alpine sandbox, then forwarded via SSH allowlist to the host.

05
LEARN

Learn

Outcome is embedded back into ChromaDB. Heartbeat pushed to the fleet cloud. Loop repeats.

Built for autonomy, not assistance

Every layer of YantraOS exists to let the machine operate without human prompting, while keeping every action auditable and reversible.

Sandboxed execution

LLM output never runs on the host. Every command routes through a locked Alpine Docker container, then through an SSH allowlist. The blast radius is always contained.

LLM output → JSON Schema validation
Docker Container (Alpine)
SSH key (allowlist only)
Host executor · exit 0

The TUI

A 3-pane Textual HUD streaming the daemon's cognitive process in real time. Telemetry, ThoughtStream, and Command, over a UNIX socket.

TELEMETRY THOUGHTSTREAM COMMAND
CPU: 12% [SENSE] read.. > _
RAM: 4.1GB [ACT] exec done

Hybrid inference

LiteLLM routes every inference call to the optimal backend based on detected hardware: NVIDIA GPU, integrated chip, or fully offline mode.

ALPHA
Ollama · local
EDGE
Cloud API
DARK
Phi-3 offline

Vector memory

ChromaDB stores every cycle outcome as a semantic embedding. The daemon retrieves relevant past context at each Kriya tick. It genuinely remembers.

Autonomous OTA

BTRFS snapshot, pacman update, systemd hook: the OS evolves itself. Every update is pre-snapshotted and reversible.

1. BTRFS snapshot →
2. pacman -Syu →
3. yantra.service restart

An AI that can execute commands is dangerous. We took that seriously.

The execution chain between the LLM and your host has four hard layers. Nothing bypasses it. No exceptions. No overrides.

Execution chain
LLM output
JSON Schema validation
Docker Container (Alpine)
SSH allowlist key
Host command executor
Threat surface analysis
[X]LLM hallucination → rm -rf /
[✓]LLM output is never executed on the host.
[X]Daemon privilege escalation
[✓]Daemon runs as yantra_daemon (UID 999).
[X]Container escape to host filesystem
[✓]Docker has no network or host mount.
[X]Unrestricted SSH execution
[✓]SSH key allows only whitelisted commands.
[X]Secrets exfiltration via prompt injection
[✓]Secrets live in /etc/yantra/secrets.env (0400). Never injected into model context.
Arch Linux·Python 3.12·asyncio·LiteLLM·Ollama·ChromaDB·Pinecone·Docker·FastAPI·Textual·Next.js·Supabase·Vercel·BTRFS·systemd·OpenSSH·Arch Linux·Python 3.12·asyncio·LiteLLM·Ollama·ChromaDB·Pinecone·Docker·FastAPI·Textual·Next.js·Supabase·Vercel·BTRFS·systemd·OpenSSH·

Your machine has been passive for too long.

YantraOS is open source and built in public. Join the fleet, contribute to the Kriya Loop, or reach out to the team.

For enterprise enquiries and partnerships:

info@euryaleferox.co