Realtime multimodal interaction
Speech-to-speech conversation, interruption handling, typed prompts, animated states, and a visual artifact surface share one desktop experience.
Working project · Local AI agent
Jarvis — AI Personal Assistant
A desktop AI companion designed to move from conversation to grounded work: voice, local files, visual artifacts, web research, images, notes, and controlled tool execution in one interface.
Product thesis
A useful personal assistant should understand the task context, show its work, and ask before crossing meaningful boundaries.
What it demonstrates
Product thinking across interface, orchestration, context, and safety.
Workspace-grounded assistance
Users explicitly select files or one active folder. Jarvis indexes supported content progressively and retrieves focused excerpts for follow-up work.
Artifacts instead of chat-only output
Markdown, Mermaid diagrams, notes, images, file previews, project trees, diffs, spreadsheets, and progress states appear in a reviewable panel.
Image and knowledge workflows
A numbered image board supports generation and editing, while xAI or Exa search can bring current web context into the workflow when configured.
Tool orchestration
Structured tool calls connect prompts to notes, workspace search, generated files, coding-assistant handoffs, and opt-in macOS computer actions.
Human-controlled execution
Sending, deleting, purchasing, changing accounts or settings, external coding runs, and private-data exposure pause for explicit confirmation.
How the workflow operates
- 01 · Intent: voice or typed input becomes a structured request.
- 02 · Context: selected files, workspace excerpts, or web results ground the task.
- 03 · Tool: Jarvis routes the request to the appropriate capability.
- 04 · Review: results appear as visible artifacts, diffs, files, or progress.
- 05 · Control: sensitive or consequential actions require explicit approval.
Build status
A working local desktop build with cloud-dependent features.
The Electron application, workspace tools, interface, TypeScript checks, production build, and dependency audit have been exercised locally. Realtime voice, image generation, and web search require the user's configured API access and available provider quota. It is presented as a desktop project—not a hosted web demo.
ElectronReactViteTypeScriptOpenAI RealtimeWebRTCxAI / ExaMermaidPDFKitExcelJS