Agent-01 is a Windows-first AI agent runtime designed to be 100% offline and privacy-focused, empowering your personal computer with Large Language Models (LLMs).

Version: 1.1.1 · Python 3.10+ · Windows-first · Single launcher: run.bat

This is not just a software application; it's a modular toy designed for the AI era. Agent-01 was 100% generated, designed, and maintained by AI. The codebase is deliberately structured to be "AI-friendly." I encourage you to open this project in modern AI IDEs (like VSCode + Antigravity, Cursor, Windsurf, etc.) and ask the AI to read, understand, expand, or completely rewrite any module you want. You are holding a framework—a skeleton. It is up to you and your AI to give it flesh and blood.

Tested and optimized for mid-range gaming and productivity rigs:

• OS: Windows 11
• CPU: Intel Core i7 10th Gen (or AMD equivalent)
• RAM: 32 GB
• GPU: NVIDIA RTX 3070 Ti (or equivalent with sufficient VRAM for GGUF models)
• Storage: Fast NVMe SSD recommended

1. Install Python 3.10+ (Node.js optional).
2. Download llama-server.exe from: https://github.com/ggml-org/llama.cpp/releases (I used: https://github.com/ggml-org/llama.cpp/releases/download/b8070/llama-b8070-bin-win-cuda-13.1-x64.zip), then extract llama-server.exe into the project's llama-server directory.
3. Configure the environment following Tutorial (EN) or Tutorial (VI).
4. Double-click run.bat.

Note: run.bat is the single entry point. It automatically checks and installs dependencies, loads environment parameters, and launches the runtime. Tip: Put machine-specific overrides (for example MODEL_PATH) in run.local.bat to keep local paths out of git history.

graph TD
    %% Theme Settings
    classDef primary fill:#f8fafc,stroke:#3b82f6,stroke-width:2px,color:#0f172a;
    classDef success fill:#f0fdf4,stroke:#10b981,stroke-width:2px,color:#0f172a;
    classDef warning fill:#fffbeb,stroke:#f59e0b,stroke-width:2px,color:#0f172a;
    classDef core fill:#eff6ff,stroke:#2563eb,stroke-width:3px,color:#0f172a,font-weight:bold;

    %% Layer 1
    subgraph L1 [1. Interface & Environment]
        direction LR
        ui_in(["[+] User Input<br/>(run.bat)"]):::success
        ui_out(["[■] Terminal UI<br/>(Output Stream)"]):::success
    end

    %% Layer 2
    subgraph L2 [2. Core Inference Engine]
        direction TB
        core(("Agent.run()")):::core
        ctx["ContextBuilder"]:::primary
        prompt["PromptBuilder"]:::primary
        llm{"LLM Inference"}:::primary
        parser["Response Parser"]:::primary
        tools["Tool Loop"]:::warning
        registry[/"Tool Plugins"\]:::warning
    end

    %% Layer 3
    subgraph L3 [3. Workspace & Persistence]
        direction LR
        sess[("Session<br/>JSON v3")]:::warning
        sum["Summarizer<br/>(3-tier)"]:::warning
        static[("Boot Files<br/>& Skills")]:::warning
    end

    %% Data Flow Connections
    ui_in -->|Input| core
    
    core -->|1. Setup Context| ctx
    ctx -->|Read| static
    ctx -.->|System Prompt| core
    
    core -->|2. Format| prompt
    prompt -->|3. Query| llm
    llm -->|4. Response| parser
    
    parser -->|Text| ui_out
    parser -->|Tool_calls| tools
    
    tools <-->|Execute| registry
    tools -.->|Results| core
    
    core -->|Save| sess
    sess -->|Check Limit| sum
    sum -.->|Compress| sess

• Offline mode (set PROVIDER=local): Default. Runs 100% offline directly on your hardware using llama-server.exe with a GGUF model.
• Online mode (set PROVIDER=openai_compatible): Uses external cloud compute to call LLMs via APIs (e.g., OpenAI, Anthropic, Google).

• Default sandbox (SHELL_WORKSPACE_ONLY=1): The agent is locked in the ./workspace directory. Modifying files outside this directory is strictly blocked.
• Desktop control (AGENTFORGE_DESKTOP_CONTROL=1): Native desktop control tools (desktop_*) via the computer-use-agents skill are enabled by default for maximum PC capability.
• Anti-loop Guards: Strict tool execution loops (MAX_ITERATIONS, MAX_REPEATS, AGENT_TIMEOUT) are wired directly from the launcher to prevent runaways.

Explore the documentation for configuration and architectural deep dives:

• Tutorial (EN) | Hướng dẫn Sử dụng (VI)
• Blueprint Architecture (EN) | Kiến trúc Hệ thống (VI)

python -m pytest -q
python -m compileall -q src
$env:PYTHONPATH='src'; python -m agentforge.cli --help

• 1.2 - Complete Photoshop utilization capabilities.
• 1.3 - Complete session management features.
• 1.4 - Integrate external APIs and access tokens (Google, Telegram, Zalo, Facebook, Discord, etc.)
• 2.0 - dual online/offline configuration.
• 2.1 - Upgrade long-term memory.
• 2.2 - Full automation via chat applications supporting chatbot creation.

Agent-01 would not exist without the following tools, platforms, and AI systems:

• Design Pattern References: PicoClaw (Go) & OpenClaw (TypeScript).
• AI Development Partners: Google Gemini 3 & 3.1, Anthropic Claude Opus 4.6, OpenAI Codex 5.3.
• Development Environment: Visual Studio Code, Antigravity by Google DeepMind.
• Third-Party Modules: adb-mcp by Mike Chambers (MIT License), llama.cpp by Georgi Gerganov, and open-source models by Meta (Llama).
• Open-Source Community: The broader open-source AI agent community.

A completely modular project, 100% generated and maintained by AI.