Frequently asked
questions.

Yes, the core JARVIS software is 100% free and open-source under the GNU General Public License v3.0 (GPLv3).

The system can run entirely on local Ollama or LM Studio models for zero API cost. If you use cloud APIs like Groq, Gemini, or Regolo, you may incur standard API usage costs depending on their respective free tiers.

Absolutely. You can point JARVIS to any local OpenAI-compatible endpoint. By running Ollama or LM Studio locally and updating PRIMARY_BASE_URL and FALLBACK_BASE_URL in your .env, you achieve 100% local, offline inference.

The local embedding model (BAAI/bge-small-en-v1.5) and RapidOCR already run on-device — no cloud calls there either.

Very little. JARVIS was developed and tested on an Intel i5-3210M (2012) with 4 GB RAM.

Because JARVIS uses cloud LLM providers by default, you don't need a powerful GPU or 32 GB of RAM. If your machine can run Windows 11 and a browser, it can run JARVIS. More RAM = faster embedding loads and snappier boot.

The AgenticBrain loop is interruptible. If you change your mind mid-task — for example, saying "not WhatsApp, send an email instead" — JARVIS absorbs the correction, revises its plan in-place, and keeps already-completed steps intact.

No full restart, no hallucinated apology loop, no lost progress. This runs only in AgenticBrain mode — FastBrain is stateless and doesn't support mid-run re-planning.

When the agent hits a missing critical parameter mid-mission — an unknown contact number, an ambiguous file target, or a design choice — it pauses on the current step and summons the glassmorphic Input Popup in Clarification Mode. The agent's exact question appears in a cyan-accented banner, background voice/STT is muted, the loop timeout clock is paused, and your typed answer is injected directly into the step's observation and confirmed facts without restarting the loop.

It's the "ask, don't guess" safety net. Once you answer, the loop resumes exactly where it left off. Only available in AgenticBrain mode.

JARVIS uses a CLA to ensure the project can remain open-source while allowing commercial licensing for enterprises.

By submitting a Pull Request, you agree to grant the maintainer (Kaif Ansari) the right to use your contribution under both GPLv3 and any future commercial licenses. This keeps the project legally clean and fully funded.

Currently, JARVIS is Windows-only (Windows 10/11). The GUI automation, PC monitor, and mobile bridge assume Windows paths and APIs (pywin32, comtypes, Windows Registry).

Linux and macOS are not currently planned, but community contributions are welcome if you'd like to port it.

Mobile control uses a persistent WebSocket bridge over local WiFi — no ADB, no USB cable, no cloud relay.

Install Termux + Termux:API from F-Droid, open the Unified Portal at http://127.0.0.1:8000, click the Mobile Bridge tab, scan the QR code, and you're connected. JARVIS can then control flashlight, vibrate, battery, camera, notifications, and open apps via termux-* API commands.

Behind the portal, the bridge still runs on 127.0.0.1:8090 — you just never need to type that URL yourself.

JARVIS employs strict shlex tokenization and path canonicalization that auto-blocks destructive commands like rm -rf / or del /f /s /q on system paths.

Additionally, the Human-in-the-Loop (HITL) system ensures it asks for consent before any major action. You see the exact operation, approve or reject, and nothing happens without your explicit confirmation.

WhatsApp uses the Baileys Node.js WebSocket bridge for stable connections, requiring a QR code scan once. Telegram uses the native MTProto API via Telethon for deeper OS integration and remote bot control.

Both are managed safely as background services by the ServiceWatchdog, which auto-restarts them if they crash.

MCP (Model Context Protocol) lets JARVIS connect to 18,000+ community-built tool servers — GitHub, Brave Search, SQLite, Notion, Slack, Docker, and more — using a single terminal command like jarvis mcp add <name> npx <package> or via Data/mcp_servers.json.

Servers hot-reload live within 2 seconds without restarting JARVIS, and their tools are automatically namespaced as mcp__<server>__<tool>.

Yes. Credentials are AES-encrypted locally. Conversation history and memory graphs stay on disk. There's zero telemetry and no analytics — you can audit the entire codebase on GitHub.

Only LLM API calls leave your machine, and you can switch to local models entirely to eliminate even that. See the full Privacy Policy.

On a warm connection, the full install takes about 5–10 minutes: the 1-click installer auto-installs Git/Python/Node via winget, clones the repo, sets up Python and Node environments, and registers the CLI.

Follow the Quick Start guide for a copy-paste walkthrough.

A single PowerShell command cleans everything. Open PowerShell and run:

Set-ExecutionPolicy Bypass -Scope Process -Force; [Net.ServicePointManager]::SecurityProtocol = [Net.SecurityProtocolType]::Tls12; & ([scriptblock]::Create((Invoke-RestMethod -Uri 'https://github.com/thekaifansari01/Jarvis-OS-Agent/releases/latest/download/uninstall.ps1')))

The uninstaller automatically:

• Kills all background processes (including the jarvis.lock file)
• Deletes JARVIS registry keys from Windows
• Removes JARVIS entries from User PATH
• Deletes global jarvis CLI shims (.cmd, .bat, .ps1)
• Deletes the install folder — .venv, node_modules, codebase and all
• Optional: HuggingFace embedding cache (~130 MB) and Documents\Jarvis media

Backup prompt (default Yes): Before deleting, it offers to back up .env API keys, Lifelong Memory Graph, indexer data, and MCP config to Desktop\Jarvis_Backup_yyyyMMdd_HHmmss\. Press Enter to accept the backup, or n to skip (permanent delete).

Full step-by-step guide: Uninstall documentation.