Yuki Ueda日本語

Yuki Ueda / Works

Personal AI agent

My own personal AI agent that handles scheduling, work, budget and travel. Talk to it from the iPhone or Mac app, and the agent running on a VPS does the research, makes the decisions and asks for approval right before anything goes out. Below is a working demo that looks just like the app.

One agent per role

A separate personal AI agent for each role. Each keeps its own memory and ledger, and acts on its own at set times (checking the schedule, scanning for gigs, and so on).

Approval right before acting

Sending, paying, booking and applying stop right before the final step and show an approval card. When a choice is needed, a choice card appears in place of the input box.

Connected by memory

Decisions, preferences and progress are written to memory as they happen. Even when a conversation is split, the next one picks up from memory.

Architecture

The app is only a window for talking and approving. The thinking and acting happen in the personal AI agent on the VPS, and anything that would reach the outside world is stopped by a fence that asks the app for approval.

Architecture of the personal AI agent: the iPhone and Mac app, one agent per role on the VPS, the fence (permission table), and the external services they connect to

Tech stack

The agent’s brainClaude Code · Claude Agent SDKSeparate conversations, memory and task ledger for each role
The owner’s appSwiftUI (iPhone, Mac)Approve on the spot from notifications and Lock Screen cards
Tools and safety fencePythonA fence that stops before sending or paying, plus scheduled jobs
Browser controlChrome DevTools ProtocolDrives a logged-in browser the way a person would
Connected servicesGoogle Workspace · GitHubGmail, Calendar, Drive, repositories
Where it runsLinux VPSRuns 24/7

← Yuki Ueda