Your personal Jarvis — running entirely on your own hardware.
A single-user AI agent with a local brain, voice, native vision, durable memory, and a living interface called the orb. No cloud required. It can even rewrite and ship its own code.
The whole console is a living green orb that is the agent. Tap to talk, drag it anywhere; the page is its canvas.
A local Qwen model on vLLM — multimodal, fast, OpenAI-compatible. Your data never leaves the box. Point at a cloud model if you have no GPU.
Continuous speech with barge-in and streaming TTS — GPU speech-to-text and neural voice, all on-device.
Share your camera and the agent sees what you show it — straight through the multimodal brain.
Charts, tables, media, 3D models, code, calendars — and bespoke web apps the agent writes and runs as floating cards.
Search, music, news, and your cloud accounts — Google & Microsoft — light up mail, calendar, and files.
Durable file memory plus semantic recall and a relationship graph. It remembers across turns and sessions.
Reach it from Telegram or WhatsApp — your own account, the full agent answering.
It can edit its own source, sandbox-test it, and promote the new version to itself — with automatic rollback.
Everything on one NVIDIA box (e.g. a DGX Spark). Most private, no recurring cost.
No GPU? Run the app locally and point the brain at a remote model — works on a Mac or any laptop.
A Windows PC with an NVIDIA card runs the whole stack via WSL2 + Docker Desktop.
Full guides: DEPLOYMENT.md · Architecture · README