ai-human

OpenAI Computer Use costs money, requires internet, and can't learn. AI Human runs free, offline, and gets smarter every day.

⚡ What AI Human Can Do Right Now

Tell it a goal in plain English. It does the rest.

"Check my emails, summarize the important ones, and draft replies for anything urgent."

"Research competitors in my market, write a report, and save it to my Desktop."

"Find all Python files modified this week, run the tests, and commit the passing ones to GitHub."

"Monitor my screen and alert me if my build breaks."

"Every Monday at 9am: generate a weekly report, email it to my team, and archive last week's."

It sees your screen, clicks, types, opens apps, writes code, reads files, sends emails, browses the web, manages databases — everything a human assistant would do.

🧠 The Capabilities That Have Never Existed Together Before

1. Self-Modifying AI — It Rewrites Its Own Code

Tell it to customize itself. It reads its own source files, modifies the code, validates with AST, takes a snapshot, restarts, and auto-rolls back if something breaks. No human in the loop.

2. Self-Creating Tools — It Builds Its Own Capabilities

Encounter a task no existing tool handles? The agent generates a new Python tool on the fly, security-validates it, hot-reloads it, and uses it — all in the same session.

3. Self-Healing — It Never Gives Up

On failure: diagnoses the error → searches the web for solutions → applies the fix → retries. It keeps trying until it succeeds or escalates to you.

4. Runs on 4GB RAM — True Democratization of AI

Auto-detects your hardware. Selects the best model that fits. TinyLlama 1.1B on 4GB. Mistral 7B on 16GB. No manual configuration needed.

5. Fully Offline — Your Data Never Leaves Your Machine

All computation happens locally via Ollama. No API keys. No subscriptions. No cloud. Your files, emails, and actions stay on your computer.

6. Learns While You Sleep

When idle for 5+ minutes, it autonomously researches knowledge gaps, analyzes past failures, extracts lessons, and stores everything in its permanent vector memory. It wakes up smarter than when it went idle.

7. Peer Network — A Swarm of AI Workers

Run AI Human on multiple machines. They discover each other via mDNS, share learned knowledge, and delegate tasks. One agent handles email; another writes code; another monitors the market.

🚀 Quick Start

Prerequisites

• Python 3.10+
• Windows 10 or 11
• 4GB RAM minimum (8GB+ recommended)
• Ollama for offline use (optional if using Claude/OpenAI API)

Install

git clone https://github.com/yourusername/ai-human.git
cd ai-human
pip install -r requirements.txt
cp .env.example .env

Configure (choose one)

Option A — Free, Fully Offline (no API key needed):

LLM_PROVIDER=ollama
# LLM_MODEL is auto-selected based on your hardware

Then install Ollama from https://ollama.com and run ollama serve. AI Human pulls the right model automatically.

Option B — Claude API:

LLM_PROVIDER=anthropic
ANTHROPIC_API_KEY=sk-ant-...

Option C — OpenAI:

LLM_PROVIDER=openai
OPENAI_API_KEY=sk-...

Run

python launcher.py                                       # recommended (auto-restart + watchdog)
python launcher.py --goal "Check my emails and summarize"  # with a starting goal
python main.py --check                                   # hardware report (what model will it use?)
python main.py --no-ui                                   # headless terminal mode

Open http://localhost:8080 in your browser for the live dashboard.

🔌 LLM Providers

Works with every major LLM provider. Switch anytime by changing one line in .env.

💻 Hardware Auto-Detection

AI Human detects your hardware and selects the best model automatically. No configuration needed.

Run python main.py --check to see your hardware tier and what will be used.

🛠️ Built-in Tools (100+)

Core Computer Control

• Mouse & keyboard: click, type, drag, scroll, hotkeys
• Window management: open, close, focus, resize any app
• File system: read, write, search, organize, bulk-rename
• Terminal: execute commands securely (no shell injection)
• Screen capture: real-time perception + change detection

Web & Research

• Web search: DuckDuckGo (no API key needed)
• Browser automation: Playwright + Chrome/Edge extension
• Web scraping: fetch and clean any URL content
• Research pipeline: search → fetch → LLM summarize → memory

Productivity

• Email: read, send, search (Gmail, Outlook, IMAP/SMTP)
• Calendar: Google Calendar integration
• Documents: read PDF, Excel, Word, PowerPoint
• Git: clone, commit, push, PR creation via GitHub CLI
• Databases: SQLite, PostgreSQL, MySQL queries

AI & Generation

• Image generation: DALL-E → Stable Diffusion → ComfyUI → Pollinations.ai (free fallback)
• Ask any AI: query ChatGPT via browser when needed
• Code execution: sandboxed Python subprocess runner
• Multi-language: 100+ languages via Google Translate / Argos (offline)

System & Automation

• PowerShell: run scripts and commands
• Windows Registry: read/write with safety guards
• Windows Task Scheduler: create and manage scheduled tasks
• Services: list, start, stop Windows services
• Process manager: list, kill, start processes

Reverse Engineering

• Binary analysis: Radare2 disassembly, function detection
• Decompiler: Ghidra → RetDec → r2ghidra auto-select
• String extraction: URLs, IPs, API endpoints, credentials
• RE report: full Markdown + JSON analysis with threat indicators

Skill Packs (Auto-Loaded)

Excel · VS Code · Windows Explorer · PowerShell · Email · Database · Clipboard · Archives · Network · Task Manager · Registry · Scheduler · Text Processing

🏗️ Architecture

launcher.py          ← Watchdog: restarts on crash, injects version, handles self-updates
  main.py            ← Entry point: wires all subsystems via AgentWiring
    core/
      agent.py       ← AgentOrchestrator: perceive → think → act → learn loop
      wiring.py      ← Clean public API for attaching subsystems
      event_bus.py   ← Thread-safe publish/subscribe between all modules
      nl_scheduler.py         ← Natural language → ParsedSchedule
      learning_loop.py        ← Autonomous idle-time research
      self_corrector.py       ← Failure diagnosis and recovery
      self_updater.py         ← Self-modification with snapshot + rollback
      peer_network.py         ← mDNS peer discovery and knowledge sharing
      credential_manager.py   ← Fernet AES-128-CBC encrypted vault
      performance_dashboard.py← JSONL metrics + Markdown reports
      auto_documentation.py   ← Background: AGENT_LOG.md, KNOWLEDGE_BASE.md
    llm/
      base.py        ← LLMProvider ABC
      factory.py     ← Hardware detection + auto model selection
      anthropic_provider.py / openai_provider.py / ollama_provider.py / lmstudio_provider.py
    memory/
      vector_store.py  ← ChromaDB PersistentClient (unlimited)
      episodic.py      ← Every action + outcome stored forever
      semantic.py      ← Learned facts and knowledge
    perception/
      screen_capture.py  ← Fast multi-monitor screenshots (mss)
      vision_analyzer.py ← Screenshot → Vision LLM → natural language
      uia_detector.py    ← Windows UI Automation (exact element positions)
      ocr_engine.py      ← Tesseract fallback
      screen_diff.py     ← Detect what changed after an action
      stress_detector.py ← Typing errors + mouse jitter → 5-level stress score
      screen_recorder.py ← MP4 recording of all agent actions
    actions/
      executor.py    ← Mouse, keyboard, file, process (no shell=True)
    tools/
      registry.py    ← Auto-discovers tools from built_in/, skills/, registry/
      built_in/      ← Core tools (always available)
      registry/      ← AI-created tools (hot-reload)
    skills/          ← Skill packs (auto-loaded)
    safety/
      broker.py          ← Risk scoring 0-1