Understudy

lastmod 2026-03-27created 2026-03-271 edits

A local desktop agent that learns from demonstrations

About

• A desktop automation agent that, once the user shows it a task, learns the intent and procedure and performs it repeatedly
• Learns directly from demonstrations without a separate workflow builder
• Controls GUI, browser, shell, file system, and messaging in a unified way from a local runtime
• 5-stage growth structure : observe → imitate → remember → optimize → autonomy
  • Layers 1–2 are implemented and usable now, 3–4 are partially implemented, 5 is a long-term goal
• The /teach command records, analyzes, optionally verifies, and publishes demonstrations
  • Intent-based learning, not simple macros
  • Results are mainly saved as SKILL.md and can be reused
• An implicit learning loop conservatively identifies repeated patterns and compresses workflows
  • Once enough repetitions accumulate, they are promoted to a workspace skill; uncertain patterns are held back to maintain safety
• Route optimization is partially implemented by preferring already-known faster routes
• Multi-channel gateway
  • Supports 8 built-in messaging channels, including Web, Telegram, Discord, Slack, WhatsApp, Signal, LINE, and iMessage
  • Terminal, WebChat, and messaging channels connect through a single HTTP/WebSocket gateway
• Currently macOS-focused; Linux and Windows GUI backends are planned

See also

• Ai agent

Favorite site

• Github - understudy-ai/understudy
• understudy - A local desktop agent that learns from demonstrations | GeekNews