Minimal by design.
Capable by evolution.
~3K lines of seed code, 9 atomic tools and a ~100-line agent loop give any LLM system-level control of a local computer. Everything else — skills, tools, playbooks — is grown at runtime and written into memory.
# ~100 lines, agent_loop.py
while not done:
obs = perceive(env)
plan = reason(llm, memory, obs)
done = execute(tools, plan)
memory.write(experience)A minimal seed with strong execution.
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, GPT, Gemini, Kimi, MiniMax, GLM and other major models. 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
Memory crystallizes throughout execution, letting the agent build stable, efficient working patterns over time.
code_runrun any codefile_readread filesfile_writewrite filesfile_patchpatch filesweb_scanperceive the webweb_execute_jsdrive the browserask_userhuman in the loopupdate_working_checkpointworking notepadstart_long_term_updatedistill long-term memoryVia code_run the agent installs packages, writes scripts and calls APIs at runtime — crystallizing temporary abilities into permanent tools.
Say it once, learn it for life.
This is what fundamentally sets GenericAgent apart from other agent frameworks.
New task
A request the agent has never seen.
Autonomous exploration
Install dependencies, write scripts, debug, verify.
Crystallize into a Skill
The execution path is written into the memory layer.
Recall next time
One sentence is enough.
| What you say | First time | Every time after |
|---|---|---|
| “Read my WeChat messages” | install deps → reverse the DB → write a reader → save Skill | one-line call |
| “Give me a morning digest of Hacker News” | write scraper → build digest → schedule daily run → save Skill | one-line call |
| “Monitor stocks and alert me” | install mootdx → build screener → configure cron → save Skill | one-line start |
| “Send this file via Gmail” | configure OAuth → write sender → save Skill | ready to use |
After a few weeks, your instance has a skill tree no one else in the world has — all grown from 3K lines of seed code.
Lighter, cheaper, and it grows.
| Feature | GenericAgent | OpenClaw | Claude Code |
|---|---|---|---|
| Codebase | ~3K lines | ~530,000 lines | Open-sourced (large) |
| Deployment | pip install + API key | Multi-service orchestration | CLI + subscription |
| Browser control | Real browser, session preserved | Sandbox / headless | Via MCP plugin |
| OS control | Mouse/keyboard, vision, ADB | Multi-agent delegation | File + terminal |
| Self-evolution | Autonomous skill & tool growth | Plugin ecosystem | Stateless between sessions |
| Out of the box | Few core files + starter skills | Hundreds of modules | Rich CLI toolset |
Five dimensions, data-backed.
Baselines include Claude Code, OpenAI Codex and OpenClaw, evaluated on Claude Sonnet 4.6 / Opus 4.6, GPT-5.4 and MiniMax M2.7 backbones.
SOP-Bench · Lifelong AgentBench · RealFin-Benchmark
Tool Efficiency Benchmark (11 simple + 5 long-horizon)
SOP-Bench (dangerous goods) · LoCoMo · 20-skill stress test
9-round LangChain longitudinal study · 8-task cross-task web benchmark
WebCanvas · BrowseComp-ZH · 22 custom tasks
A real, persistent browser — not a headless sandbox.
GA web tools run through TMWebdriver: a local WebSocket server plus a Chrome extension driving your real Chrome session, preserving cookies, login state, extensions and normal fingerprints.
| Detection signal | Vanilla headless automation | GA web tools |
|---|---|---|
| SannySoft headless test | Often detected | 56/56 passed |
| bot.incolumitas.com | Fails webdriver / CDP checks | 36/36 passed |
| BrowserScan bot detection | Often abnormal | Normal |
| Device & Browser Info | Multiple bot flags | Human · isBot=false |
| FingerprintJS bot demo | Often detected | Passed |
| reCAPTCHA v3 demo | Low human score | 0.9 human-like score |
Use it from a terminal, a desktop app, or your favorite chat.
Terminal UI
prompt_toolkit + rich, multiple sessions, streaming. python frontends/tui_v3.py
Desktop & web UI
A ready-to-run desktop app ships with one-line installs; developers use python launch.pyw.
IM bots
Telegram, Discord, Lark/Feishu, WeChat, QQ, WeCom and DingTalk frontends.
Android
The same agent inside your phone — see GA Mobile.
GA MobileGoal mode
Time-budgeted self-driven loop: “keep optimizing X for N hours” without premature delivery.
Goal Hive
BBS-coordinated master/workers running long-horizon objectives in parallel.
Conductor
Spawn, supervise and auto-clean parallel sub-agents.
Reflect mode
Drive GA from an external tick — how local agents join GA Arena with a single bridge file.
GA ArenaRead the code in an afternoon.
MIT licensed. Star it, fork it, or hand it your first task.