Introduction

Personal Jarvis is a voice-driven meta-orchestrator. You speak; it routes the work to whatever tool can actually do the job — Claude Code, the Codex CLI, Open Interpreter, MCP servers, or your computer directly — and reports back. The voice layer is just the interface. The product is the dispatcher behind it.

Not another chatbot

Most “voice assistants” are a single language model with a microphone bolted on. Jarvis is the opposite: a Supervisor-Agent that classifies every spoken command and dispatches it to an interchangeable harness. Small talk stays local and instant. Real work — “open a PR”, “fix that failing test”, “install this CLI” — is handed to a background worker that runs in isolation and verifies its own output.

The core UX contract is one uninterrupted spoken conversation. Jarvis acknowledges you in under a second, then does the heavy lifting off the transcript while you keep talking.

The three things that make it different

• Harness-agnostic. Every backend — Claude Code, Codex, Open Interpreter, MCP, computer-use — sits behind one protocol. Adding a new engine never touches the voice layer.
• Self-healing. Complex tasks run as Missions in an isolated git worktree. A Critic reviews the result and retries up to three times before it ever speaks back to you.
• Yours, end to end. It runs on your machine, with your keys, in your credential manager. No Jarvis server sees your data, your prompts, or your keys.

Who it is for

Developers and power users who want a hands-free way to drive the AI tools they already use without surrendering their keys or their data to a Jarvis-operated hosted service. Windows is the hardware-tested desktop target; macOS and Linux code paths remain unverified on real desktop hardware.

Next: install it in one line, then hand over your keys on first run.