What it does GenericAgent gives an LLM direct control over your local machine—browser, terminal, filesystem, keyboard, mouse, screen vision, even mobile devices via ADB—through just nine atomic tools and a roughly 100-line execution loop. The entire core fits in about 3,000 lines of Python. It solves tasks by exploring, then “crystallizes” the working solution into a reusable Skill stored in layered memory. Use it long enough and you end up with a unique personal skill tree no one else has.

The interesting bit The author claims they never opened a terminal during development: GenericAgent allegedly ran git init, installed dependencies, and wrote every commit message autonomously. Whether that’s fully accurate or a neat demo, the design philosophy is the hook—don’t ship a bloated framework, ship a seed and let the agent grow its own environment.

Key highlights

• Minimal footprint: ~3K lines core, ~100-line agent loop, <30K token context window (vs. 200K–1M for comparable agents)
• Self-evolving skills: First time you ask it to read WeChat messages it reverse-engineers the DB and writes a script; next time it’s a one-liner
• Real browser injection: Preserves login sessions, not a sandboxed headless browser
• Multi-model support: Works with Claude, Gemini, Kimi, MiniMax, and others
• Multiple frontends: Desktop app, Textual TUI, Streamlit, plus IM bots for Telegram, WeChat, QQ, Lark, WeCom, DingTalk

Caveats

• Windows TUI rendering is explicitly noted as “flaky” depending on terminal and font; Git Bash recommended over PowerShell/cmd
• Python 3.14 is incompatible due to pywebview issues; stick to 3.11 or 3.12
• One-line installer pulls from http://fudankw.cn:9000—not a standard package registry

Verdict Worth a look if you’re exhausted by 500K-line agent frameworks and want something you can actually read in an afternoon. Skip it if you need battle-tested enterprise reliability or don’t trust an agent with system-level mouse and keyboard control.