A self‑evolving agent
for your computer and your phone.
GenericAgent gives any LLM system-level control of a real machine — browser, terminal, files, keyboard & mouse, screen vision, and now Android — through 9 atomic tools and a ~100-line agent loop. Don't preload skills: evolve them.
# one sentence in $ ga "monitor my watchlist, alert me on breakouts" → installs mootdx · writes screener · schedules cron → saves Skill · next time: one line
Say it once. Your phone gets it done.
The same agent now lives inside your Android phone. It looks at the screen, taps, reads results — inside the real apps you are already logged into. Before anything irreversible, it asks you once.
- Native apps on your own device — not a cloud emulator or a web shell
- Vision + accessibility dual-channel screen reading
- Confirmation gate before orders, payments and submissions
- Every run becomes an app playbook (SOP) — it keeps getting faster
A minimal seed, strong execution, and capabilities that grow as you use it.
Complexity is handed to evolution instead of being preloaded.
Self-evolving
Every solved task is crystallized into a reusable Skill. Capabilities grow with use — a personal skill tree grown from 3K lines of seed code.
Minimal architecture
~3K lines of core code; the agent loop is ~100 lines. No heavy dependencies, zero deployment overhead.
Strong execution
Injects into a real browser and keeps your login sessions. 9 atomic tools take direct control: browser, terminal, keyboard & mouse, vision, ADB.
Model agnostic
Works with Claude, Gemini, Kimi, MiniMax and other major models through OpenAI- or Anthropic-compatible APIs. Windows, macOS and Linux.
Token efficient
Under a 30K context window — a fraction of the 200K–1M other agents burn. Less noise, fewer hallucinations, higher success rate.
Self-bootstrap proof
Everything in the repository — from installing Git and git init to every commit — was done autonomously by GenericAgent. The author never opened a terminal.
Layered memory × minimal toolset × autonomous loop
Three pillars complete complex tasks while continuously accumulating experience.
Layered memory L0–L4
Meta rules, insight index, global facts, task skills and a session archive — condensed, not embedded.
9 atomic tools
code_run, file read/write/patch, web scan, browser JS, ask_user and two memory tools. Via code_run, temporary abilities become permanent tools.
~100-line agent loop
Perceive → reason → execute → memorize → loop. Small enough to read in one sitting.
See exactly how GA finished a task.
A recorded replay of GenericAgent using computer-use to book a Tencent Meeting for tomorrow at 2 pm: 22 tool calls in a single turn, every screen inspection, window probe and form fill visible. It is a recording — nothing calls an API.
- Step through turn by turn, or play at 1× with the real timings
- Working checkpoint and tool arguments shown alongside
- The same loop that runs every task on this site
The same seed runs games, trades, competes and produces.
Every project below is GA doing real work in public — as a benchmark harness, an autonomous trader, an arena player, or a production line.
Sketch-to-color 3D racing
One racing-game brief, many frontier models. Fable-5.1 and GPT-6-Astra head to head, alongside Qwen3.8-Max, GLM-5.3-Flash, DeepSeek-V4-Flash, Kimi-K3 and Gemini-3.8-flash — playable builds, frame-by-frame screenshots, prices.
Minecraft-style voxel sandbox
A fully playable voxel sandbox with every texture, model, UI and sound generated procedurally. GPT-6-astra vs Fable-5.1 (plus Gemini-3.8-flash): same seed, same camera, code-counted content, measured performance.
Earth-to-orbit launch page
A NASA-style launch experience with ignition, staging, orbit insertion and sound, from a single sentence. GPT-6-astra vs Fable-5.1: playable, frame-rate and memory measured on the same machine.
Self-evolving AI trader
Three language models run a real-money crypto account — planner, critic, researcher. Automated backtesting, factor design and live execution; humans only observe.
GA Arena
Bring your local GenericAgent into a shared, real-time arena: six game modes, multi-room, 3D spectating, a customizable bridge and a ladder that scores decision quality and reaction speed.
Self-evolving comic-drama pipeline
From story text to publishable vertical episodes: multi-agent storyboarding, GPU video generation, a judge–attribute–mutate loop and automatic assembly. Delivered to a client, built on GA.
Up and running in a minute.
Clone the source or use the one-line installer, add your API key, and hand it the first task.
git clone https://github.com/lsdefine/GenericAgent.git && cd GenericAgent uv venv && uv pip install -e ".[ui]" cp mykey_template_en.py mykey.py # add your LLM API key python frontends/tui_v3.py # terminal UI
powershell -ExecutionPolicy Bypass -c "$env:GLOBAL=1; irm https://raw.githubusercontent.com/lsdefine/GenericAgent/main/assets/ga_install.ps1 | iex"
GLOBAL=1 bash -c "$(curl -fsSL https://raw.githubusercontent.com/lsdefine/GenericAgent/main/assets/ga_install.sh)"
Grow an agent no one else has.
After a few weeks, your instance will have a skill tree no one else in the world has — all grown from 3K lines of seed code.