GenericAgent

// Overview

GenericAgent is a minimal autonomous agent framework whose core is about 3,000 lines of code, built around nine atomic tools and an agent loop of roughly 100 lines. Those tools give a model system-level control of a local computer: a real browser, the terminal, the filesystem, keyboard and mouse input, screen vision, and Android devices over ADB. The design philosophy the README states is 'don't preload skills, evolve them.'

Every time the agent solves a new task it crystallises the execution path into a reusable Skill, so capability accumulates into a personal skill tree grown from the seed code rather than shipped with it. Browser work goes through TMWebdriver, which injects into a real browser and preserves logged-in sessions instead of driving a fresh headless instance. The project reports running under a 30K context window — a fraction of what agents in the 200K–1M range consume — which the authors tie to less noise and lower cost per task.

Dependencies are deliberately tiered: the agent core needs only `requests` plus four lightweight packages for TMWebdriver's local server, with Streamlit and the TUI libraries behind an optional `[ui]` extra — no Playwright, no LangChain, no browser binaries to download. It works with Claude, Gemini, Kimi, MiniMax and other major models and runs cross-platform. The code is MIT-licensed; a technical report accompanies it on arXiv, and the maintainers note that the GitHub repository and gaagent.ai are the only official channels.

// What it does

• Nine atomic tools covering browser, terminal, filesystem, keyboard/mouse, screen vision and mobile (ADB)
• Self-evolution: each solved task is crystallised into a reusable Skill, building a personal skill tree
• TMWebdriver injects into a real browser and keeps existing login sessions
• Runs under a ~30K context window rather than the 200K–1M other agent frameworks consume
• ~3K lines of core code with a ~100-line agent loop — the core needs only requests plus four small packages
• Model-agnostic: Claude, Gemini, Kimi, MiniMax and other major providers
• Ships a terminal UI and a Streamlit web UI, or runs headless if you skip the [ui] extra

// Getting started

Use Python 3.11 or 3.12 — the README warns that 3.14 is incompatible with pywebview and several other dependencies. A detailed guide lives in docs/installation.md.

git clone https://github.com/lsdefine/GenericAgent.git && cd GenericAgent
uv venv && uv pip install -e ".[ui]"

cp mykey_template_en.py mykey.py
$EDITOR mykey.py

python frontends/tui_v3.py   # terminal UI
python launch.pyw            # Streamlit web UI

curl -fsSL https://raw.githubusercontent.com/lsdefine/GenericAgent/refs/heads/main/docs/installation.md

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Automate desktop and browser work that needs an existing logged-in session rather than a clean headless browser
• Build up a private, reusable skill library from tasks you actually run instead of curating one up front
• Run an agent under a tight token budget where a 200K-context harness is too expensive
• Drive an Android device over ADB from the same agent loop that drives the desktop

// How GenericAgent compares

GenericAgent alongside other open-source agent frameworks & builders tools AI/TLDR tracks, ranked by GitHub stars.