JARVIS

A voice-first AI operating layer for Windows with eight specialist agents behind one wake word.

What it is

Say a wake word and JARVIS runs your laptop. A supervisor plans the task and delegates to specialist agents (Operator, Navigator, Researcher, Engineer, Courier, Pilot, Librarian, Critic). Every action passes a permission gate before MCP servers for system, files, browser, apps, media, screen, memory and more act on the PC. An Android companion approves actions and controls the PC remotely.

How it works

1. A local wake word and streaming speech recognition start a command.
2. A model router chooses between Claude, OpenAI, Gemini and local Ollama with failover and budgets.
3. The supervisor delegates to specialist agents and tracks a task graph.
4. A permission gate checks every action; requests are signed (Ed25519) with replay protection; deletes go to the Recycle Bin.
5. Eleven first-party MCP servers do the work; skills are SKILL.md bundles it can load and write.

Engineering and proof

• About 1,180 automated test cases in 119 test files, using mocked models and tools so tests never call real APIs.
• A TypeScript monorepo of roughly 80,000 lines with an Electron HUD and an Expo plus Kotlin Android app.
• Built over about a week with parallel AI coding agents; architecture, security model and tests are documented.

My role

Designed and built by Sai Charan, with AI coding agents.

Built with

• TypeScript
• Node.js
• Electron
• React Three Fiber
• MCP
• Ollama
• Expo
• Kotlin
• Vitest

Links

Ask for a walkthrough

The repository is private. A live demo is available on request.