Open Source · GPLv3 · Live

Not a chatbot.
An autonomous OS agent.

JARVIS is an open-source AI operating system for Windows. It reads codebases, sends emails, controls your Android, remembers for months, and re-plans mid-task when you interrupt. Runs on 4GB RAM.

Star on GitHub play_circle See live demo

jarvis — agentic loop — zsh

MCP Tools

RAM

Memory

Open Source

codePython 3.10+ codeNode.js 18+ record_voice_overDeepgram Nova-2 cloudAzure Speech databaseChromaDB hubNetworkX visibilityOpenCV text_fieldsRapidOCR memoryOllama memoryLM Studio extensionMCP Protocol desktop_windowsPyQt5 boltZeroMQ check_circlepytest terminalFlask boltGroq auto_awesomeGemini graphic_eqSarvam TTS micVosk

How it works

Wake. Route. Act.

Three steps. Zero configuration. Every layer has a fallback so nothing blocks you.

Wake

Say "Jarvis" — the on-device Azure Custom Keyword model picks it up in under 200ms with 97%+ accuracy. Or hit Ctrl+Shift+J to type. No cloud round-trip either way.

Route

A semantic router decides per command: FastBrain for simple chat (<2s), AgenticBrain for multi-step work. Three tiers of fallback if the router itself fails.

Act

Real tools, real effects. Files written, emails sent, calendars updated. Interrupt mid-run — the loop re-plans per step. Nothing single-point-of-failure.

Proof of work

Three demos. Zero edits. On a 2012 laptop.

Raw, unedited console logs from an Intel i5-3210M / 4 GB RAM machine. No cloud VMs. No beefy rigs. Just one mind, working.

Demo 01

The Principal Architect

Voice command builds a production-grade Snake game — then intercepts a WhatsApp video mid-build and asks for consent before acting.

"Create advanced game UI and graphics for Snake."

Steps24

Code35K chars

First audio1.99s

Demo 02

The Autonomous Developer

Builds a SQLite CLI todo app + pytest suite. Self-corrects a Windows path bug and absorbs a mid-run feature request — without restarting.

"Arre wait, also add a priority field and make list sort by priority."

Steps7

Tests12/12 ✓

First audio0.81s

Demo 03

The Deep Planner

One vague voice command → 8 minutes of reasoning → a 12,513-char production tool with DPAPI + AES-256-GCM. Zero retries, zero errors.

"Jarvis, create a test script for my computer."

Reasoning8 min

Files2 · 1 call

Errors0

Signature feature

Interrupt it.
It adapts.

Most agents are fire-and-forget. Once a plan starts, you either wait — or kill the whole run. JARVIS listens while it works, absorbs corrections mid-flight, and keeps everything it already finished.

check_circle No full restart — completed steps stay intact.
check_circle No hallucinated apology loop — the plan graph is mutable.
check_circle Works mid-flight — corrections land in the current loop, not the next session.

Without JARVIS

            ▶ start: build + send + schedule
            ✓ step 1: project scaffold
            ✓ step 2: files written
            🚨 user: "actually, email not WhatsApp"
            ✗ KILL RUN
            ✗ RESTART FROM STEP 1
            ✗ ALL PROGRESS LOST
          

With JARVIS

            ▶ start: build + send + schedule
            ✓ step 1: project scaffold
            ✓ step 2: files written
            🚨 user: "actually, email not WhatsApp"
            🧠 re-plan: swap channel only
            ✓ email sent to corrected address
            ✓ calendar scheduled · task complete
          

Capabilities

Everything an OS agent should be.

Not a wrapper around an LLM. A full autonomous system — memory, voice, tools, routing, self-healing, and remote control in one continuous mind.

Signature

sync

Live Correction + Mid-Task Clarification

The agent loop is genuinely interruptible. Change your mind mid-task — JARVIS re-plans in-place, keeps completed steps intact, and never restarts from scratch. When critical parameters are missing, it pauses and asks you via ask_user instead of guessing.

28🚨 User injected: "add a priority field and sort list by it"
29🧠 Re-planning: install pytest → add priority column → extend tests
30❓ ask_user: "Which email should I send the summary to?" → loop paused
31✓ 12/12 tests passed — 4 new ones from the override

Memory

memory

Lifelong Memory

Five-tier memory with a weighted property graph, temporal decay, and hybrid BM25 + vector RAG.

01Episodic · 15-day rolling

02Semantic · knowledge graph

03RAG · hybrid BM25 + vector

04Ephemeral · 10m TTL state

05Behavioral · pattern engine

Voice

record_voice_over

Dual-Engine Voice

Azure Custom Keyword (on-device, 97%+) with DSP conditioning — auto-falls back to Vosk. Zero cloud round-trip.

MCP

extension

MCP-Native

Any of 18,000+ community servers — GitHub, Brave, SQLite, Notion, Docker — via a single JSON entry.

github brave sqlite notion docker slack +18k more

Mobile

phone_android

Android Control

Persistent WebSocket bridge over local WiFi. Calls, SMS, GPS, torch, clipboard — no SSH, no cloud.

Vision

visibility

Vision GUI

Set-of-Mark overlay with OpenCV edge detection. Click-by-ID with text verification, not blind pixel clicks.

Proactive

notifications_active

Proactive HITL

Background listeners for Gmail, WhatsApp, Telegram, Calendar. Distinguishes FYI from consent-requiring actions.

Hardware

developer_board

Runs on 4GB RAM

Verified on an Intel i5-3210M (2012). Lightweight local embeddings + cloud LLMs. No GPU required.

Architecture

Two brains. One mind.

A semantic router picks the right brain per command — sub-2s for simple chat, deep multi-step reasoning for real work.

bolt

FastBrain

Stateless · <2s

Short commands, casual chat, OS toggles. Fast, cheap, always on.

check_circleOpen / close apps, URLs, YouTube
check_circleVolume, brightness, mute, screenshot
check_circle10-turn rolling history + system log
check_circleWeb search + quick memory recall

psychology

AgenticBrain

Stateful · Multi-step

Complex engineering, comms, memory, vision. Full tool use, up to 50 steps.

check_circleFull file CRUD + repo-map + diffs
check_circleGmail · WhatsApp · Telegram · Calendar
check_circleLive mid-execution correction
check_circleMCP tools · vision · 5-tier memory

Under the hood

A resilient system, not a wrapper.

Every layer has a fallback. Every provider has a failover. Every tool has a watchdog. Nothing single-point-of-failure.

Input

🎙️ Voice

⌨️ Ctrl+Shift+J

📱 Telegram Bot

📡 Proactive Listeners

Routing

🧭 Semantic Router

🔁 Local Rule Fallback

Brains

⚡ FastBrain

🧠 AgenticBrain

Providers

Primary (OpenAI-compat)

Fallback (OpenAI-compat)

Ollama · LM Studio · vLLM

Tools & Memory

💻 Code

📨 Comms

🖱️ GUI

📱 Mobile

🔌 MCP

🧠 5-tier Memory

📚 Hybrid RAG

Watchdog

🛡️ ServiceWatchdog

🔄 Provider Failover

⚡ Two-Strike Recovery

Honest comparison

Where JARVIS fits.

If you need a polished IDE assistant, use Claude Code. If you want an open, hackable, cross-device autonomous OS agent — this is it.

Capability	JARVIS OS	Claude Code	AutoGPT / CrewAI
Cost model	check_circleFree, self-hosted	Paid subscription	Free / paid
Runs on 4 GB RAM	check_circleVerified	errorHeavy	errorVaries
Local LLM support	check_circleOllama, LM Studio	cancelNo	errorPartial
Model-agnostic framework	check_circleAny OpenAI-compat	cancelAnthropic only	errorPartial
Live mid-execution correction	check_circleInterruptible	cancelRestart-based	cancelNo
Line-drift-free edits	check_circleExact block diffs	errorLine-based	cancelNo
Lifelong memory	check_circleWeighted graph	cancelSession-only	errorLimited
Native real-world tools	check_circleEmail, WhatsApp, Calendar	cancelNo	errorVia plugins
Plug-and-play MCP	check_circleAny MCP server	errorLimited	cancelNo
Voice-first interaction	check_circleDual-engine wake	cancelNo	cancelNo
Remote control	check_circleTelegram + Android	cancelNo	cancelNo
Proactive background agent	check_circleHITL Scout	cancelNo	cancelNo
Multi-tier router fallback	check_circle3-tier + local	cancelNo	cancelNo
Enterprise support	errorSolo project	check_circleAnthropic	Varies

Measured, not claimed

Real numbers, real hardware.

Windows 11 · Intel i5-3210M (2012) · 4 GB RAM · CPU-only inference. Your mileage will vary by provider.

~50s

Cold boot to ready

with lazy service init

0.81s

First audio

measured, Demo 02

~810ms

Router decision

semantic tier-1

<2s

FastBrain reply

typical stateless

<100ms

Router failover

3-tier chain

~7s

Memory graph recall

weighted subgraph

~2s

Incremental RAG scan

hash-only changes

<4GB

Peak RAM footprint

shared with Windows

Get running

Up and running in five commands.

Windows 10/11 · Python 3.10+ · Node 18+. One-block setup. Zero cloud setup required for local mode.

            terminal
            PowerShell
          
# 1. Clone

git clone https://github.com/thekaifansari01/Jarvis-OS-Agent.git

cd Jarvis-OS-Agent

# 2. Virtual env

python -m venv .venv

.\.venv\Scripts\Activate.ps1

# 3. Install

pip install -r requirements.txt

pip install mcp

# 4. WhatsApp bridge

cd tools/Messanger/whatsapp/BaileysServer

npm install

cd ../../../..

# 5. Register global CLI

python SetupRegistry.py

# 6. Launch

jarvis

01 / CONFIG

Copy `.env.example`

Add API keys for your LLM provider. Works with Groq, Regolo, OpenRouter, Ollama, LM Studio, or any OpenAI-compatible endpoint.

02 / OPTIONAL

Enable MCP

Run jarvis mcp add to hot-load 18,000+ tools live — GitHub, Brave, SQLite, Notion, Docker — with zero restarts.

03 / READY

Say "Jarvis"

Wake word is on-device. Say it, and start talking. Or hit Ctrl+Shift+J for text mode.

Summon anywhere

One hotkey. Anywhere.

While JARVIS runs in the background, hit this anywhere in Windows to pull up the text input popup.

Ctrl + Shift + J

Or just say "Jarvis" — the on-device wake word listens 24/7.

Questions

Frequently asked.

Is it really free?add

Yes. GPLv3. You can run it entirely on local LLMs with zero API cost, or mix free-tier cloud providers. No subscription, no feature gate, no telemetry.

Does it work on Mac or Linux?add

Currently Windows-only. GUI automation, PC monitor, and the Termux bridge assume Windows paths and APIs. Cross-platform support is on the roadmap.

Is my data safe?add

Credentials are AES-encrypted locally. Conversation history and memory graphs stay on disk. Only LLM API calls leave your machine — and you can switch to local models entirely with zero cloud calls.

What hardware do I need?add

JARVIS was developed on an Intel i5-3210M with 4 GB RAM. If your machine can run Windows 11 and a browser, it can run JARVIS. More RAM = snappier local embeddings and faster boot.

Does RAG / memory need API keys?add

No. Both the LTM semantic graph and RAG index use the local BAAI/bge-small-en-v1.5 model (33M params, ~130 MB). Embeddings run on CPU with zero cloud calls, quotas, or cost. Only LLM/voice/comms need keys.

Which LLMs does it support?add

Any OpenAI-compatible endpoint. Configure two providers — PRIMARY_* and FALLBACK_* — and JARVIS auto-routes on failure. Works with Groq, Regolo, OpenRouter, Gemini's OpenAI endpoint, Ollama, LM Studio, vLLM, or any OpenAI-schema API.

Can I add capabilities without editing code?add

Yes — via MCP. Run jarvis mcp add <name> <command> (or save Data/mcp_servers.json), and JARVIS hot-reloads the server live within 2 seconds without restarting. Tools appear automatically as mcp__<server>__<tool>.

How is this different from ChatGPT or Claude?add

Those are chat interfaces. JARVIS is an agent — it takes actions on your actual system, remembers you across months, corrects itself mid-task, and reaches out proactively. It doesn't just answer; it does.

Your machine.
Fully autonomous.

Join the developers building an open, hackable, cross-device OS agent that respects their hardware. Star the repo, fork it, run it on your 2012 laptop.

Star on GitHub arrow_forward Read the docs

Not a chatbot. An autonomous OS agent.