anarchy-ai 1.1.1

ANARCHY - Local AI Shell with No Cloud Dependencies

ANARCHY — Local Windows AI Assistant

  ██████╗  ███╗  ██╗ █████╗ ██████╗  ██████╗ ██╗  ██╗ ██╗   ██╗
  ██╔══██╗ ████╗ ██║██╔══██╗██╔══██╗██╔════╝ ██║  ██║ ╚██╗ ██╔╝
  ███████║ ██╔██╗██║███████║██████╔╝██║      ███████║  ╚████╔╝
  ██╔══██║ ██║╚████║██╔══██║██╔══██╗██║      ██╔══██║   ╚██╔╝
  ██║  ██║ ██║ ╚███║██║  ██║██║  ██║╚██████╗ ██║  ██║    ██║
  ╚═╝  ╚═╝ ╚═╝  ╚══╝╚═╝  ╚═╝╚═╝  ╚═╝ ╚═════╝ ╚═╝  ╚═╝    ╚═╝

NO CLOUD. NO TELEMETRY. NO GODS. ONLY ZUL.

ANARCHY is a fully local, offline-first AI assistant that runs entirely on your hardware. No API keys, no subscriptions, no data leaving your machine. It understands plain English — you don't use shell commands.

---

What It Does

Capability	Description
Plain-language input	Say "write a file to my desktop that says test" — no commands, no syntax
Files & Code	Read/write files, lint, explain, generate, refactor, search codebases
Git & GitHub	Status, diff, commit, push/pull, PRs, issues — all from natural language
Desktop Automation	See the screen, click, type, scroll, drag, manage windows — 5-tier permission system
Windows UIA	Element-level control of Win32/WPF/UWP apps by name/role — more reliable than pixel clicking
Browser Control	Chrome/Edge via Playwright CDP: navigate, click, fill forms, extract data
System Settings	Dark mode, brightness, Wi-Fi, power plan, volume — natural language, no menus
Voice Input	Whisper-powered push-to-talk, wake word detection (no cloud)
Timers & Notifications	"remind me in 30 minutes" — Windows toast + voice callback
Task Scheduler	Persistent Windows scheduled tasks that survive reboots
AI & Memory	Semantic RAG search, knowledge graph, multi-LLM routing, episodic memory
Skills	Reusable multi-step workflows: code review, security audit, browser tasks, dev setup
MCP Server	Exposes memory and skills to Claude Code, Cursor, and any MCP client

---

Install in One Command

Windows (recommended)

Open PowerShell (Win + R → type powershell → Enter) and run:

irm https://raw.githubusercontent.com/psfr4590-afk/ANARCHY/main/get-anarchy.ps1 | iex

Installs Git, Python 3.11, and Ollama automatically on a fresh machine. Nothing to pre-install.

GPU? ANARCHY auto-detects NVIDIA/AMD and configures GPU acceleration.
No GPU? Runs on CPU. Works on any Windows 10/11 machine with 8 GB RAM.

macOS / Linux

curl -fsSL https://raw.githubusercontent.com/psfr4590-afk/ANARCHY/main/get-anarchy.sh | bash

Manual

git clone https://github.com/psfr4590-afk/ANARCHY.git
cd ANARCHY
py install.py        # Windows
python3 install.py   # macOS / Linux

---

Launch

# Windows
cd %USERPROFILE%\anarchy && py launch.py

# macOS / Linux
cd ~/anarchy && python3 launch.py

On launch you get:

• System tray icon — right-click to approve/deny actions, toggle voice, open dashboard, exit
• Global hotkey Win+Alt+Z — focuses ANARCHY from anywhere on your desktop
• Approval overlay — native floating card replaces CLI prompts for all state-changing actions
• ANARCHY> REPL — type anything in plain English

---

Optional Windows Features

Install these for the full native experience:

# Native approval cards + system tray
pip install pystray pillow

# Global hotkeys
pip install keyboard

# Windows toast notifications
pip install win11toast

# Element-level Windows UIA control
pip install pywinauto

# Browser automation (Playwright)
pip install playwright && python -m playwright install chromium

# Wake word detection (hands-free)
pip install openwakeword sounddevice

# Timer voice announcements
pip install pyttsx3

# Semantic routing (smarter tool matching)
pip install sentence-transformers

---

Plain Language Examples

ANARCHY understands what you mean — no shell syntax required:

write a file to my desktop that says test
show me what's in my downloads folder
analyze C:\Users\krash\Desktop\myproject
check the GPU temperature
open Chrome and go to github.com
click the OK button in the Settings window
set my screen to dark mode
remind me in 30 minutes
schedule a daily backup every night at midnight
git commit with a message about fixing the login bug
security review core/auth.py

---

Key Commands

help                    Full command reference
status                  System health + subsystem dashboard
hw                      Live hardware snapshot (CPU / GPU / RAM)
models                  List all Ollama and GGUF models

# Voice
voice on / off          Enable / disable push-to-talk
voice listen            Record one utterance and transcribe

# Timers
timer start lunch 45m   Start a 45-minute countdown named "lunch"
timer list              Show all running timers
timer cancel lunch      Cancel a timer

# Windows Settings
windows_settings        Adjust display, audio, power, Wi-Fi via natural language

# Scheduler
schedule list           Show all scheduled tasks
schedule add <n> <interval> <prompt>   Create a scheduled task

# RAG / Code Search
index [path]            Index a directory for semantic search
ask <query>             Search your codebase in plain English

# Git
commit [--all] [-m msg] Auto-draft commit message from staged diff
commit-push             Commit → push → PR in one step
hooks install-anarchy   Install ANARCHY pre-commit hook

# Skills
skill list              List installed skills
skill install <url>     Install a skill from GitHub URL

# Blueprints
blueprint list          List saved workflows
blueprint run <name>    Replay a saved workflow

# Code Intelligence (LSP)
diagnose <file>         Run language server diagnostics
diagnose                List installed language servers

# Planning
plan <goal>             Generate + execute a multi-tool plan
perf                    Performance dashboard

---

Permission System

All tool execution flows through a non-bypassable AuditGateway. Every action is checked against config/permissions.json — the single source of truth. A safety floor prevents critical tools from ever being manifested below their minimum tier.

7 tiers:

Tier	Auto-execute	Confirmation	Examples
SAFE	✓ Always	None	file_read, git_status, hardware_snapshot, timer
CONFIRM	✗	[y/N] prompt or overlay card	file_write, git_commit, windows_settings, browser
HIGH_RISK	✗	Must type "yes"	shell_exec, self_modify, kill_process
T0_OBSERVE	✓ Always	None	screenshot, list_windows, clipboard_read
T1_NAVIGATE	✓ in auto-mode	[y/N] in user-mode	mouse_move, window_focus
T2_CONTROL	✗	Always confirm	mouse_click, type_text, key_press
T3_ELEVATE	✗	Always confirm	clipboard_write, open_url, run_command
T4_OVERRIDE	✗	Double-confirm	force_close_window, system_lock, run_as_admin

Approval flow for every state-changing action:

1. Gateway classifies the action and builds ActionMetadata (target, expected effect, rollback hint, diff)
2. Native overlay card shown (if display available) with full context
3. CLI fallback if no display
4. Human decision (approved/denied) recorded in logs/security.jsonl

Unknown tools default to CONFIRM — not SAFE. ANARCHY fails closed.

---

Built-in Skills (14)

Invoke by describing what you want in plain English:

Skill	What it does
security_review	Deep audit across 7 categories: injection, auth, secrets, crypto, dependencies, error handling, input validation
code_review	6-dimension review: comment accuracy, test gaps, silent failures, type design, maintainability, simplification
audit_repo	Full repo audit: file tree, complexity hotspots, TODO search, git status
refactor_repo	Map structure, find hotspots, apply targeted edits with confirmation
generate_documentation	Auto-generate Markdown docs for all tools, skills, and workflows
backup_project	Hash, zip, and verify a full project backup
run_diagnostics	Full system diagnostic: hardware, processes, network, LLM, tool registry
summarize_directory	Human-readable summary of every file in a directory tree
open_app	Launch, focus, or close any Windows application
system_tune	Adjust display, audio, power plan, dark mode, Wi-Fi via natural language
browser_task	Multi-step browser workflows via Playwright
setup_dev_env	Clone repo → install deps → open in VS Code
daily_brief	Morning digest: hardware, health, recent activity, scheduled tasks
dockerfile_review	Security, size, and best-practice review for Dockerfiles

---

Tools (73)

ANARCHY ships 73 registered tools across 10 domains:

Files: file_read, file_write, delete_file, list_dir, search_files, copy_file, move_file, rename_file, create_directory, delete_directory, zip_files, unzip_file, hash_file, file_metadata

Code: codegen, lint_code, code_explain, code_run, code_search, code_refactor, code_test_generate

Git / GitHub: git_status, git_log, git_diff, git_commit, git_push, git_pull, git_clone, git_checkout, git_branch_list, git_stash, github_list_repos, github_get_repo, github_list_issues, github_create_issue, github_comment_issue, github_pull_request_info, github_pull_request_review, repo_map, repo_search, repo_analyze, repo_refactor_plan

Desktop (T0–T4): screen_capture, screen_size, mouse_position, pixel_color, clipboard_read, list_windows, get_active_window, find_on_screen, mouse_move, mouse_scroll, window_focus, window_minimize, window_maximize, window_resize, window_move_pos, mouse_click, mouse_double_click, mouse_drag, type_text, key_press, key_hotkey, clipboard_write, window_close, open_url, force_close_window, system_lock, system_sleep, run_as_admin, input_block

Windows-native: windows_uia, windows_settings, win_notifications, timer, task_scheduler_tool

Browser: browser (Playwright CDP — Chrome/Edge)

App adapters: vscode, windows_settings

System: hardware_snapshot, system_diagnostics, system_info, network_info, battery_info, wifi_scan, bluetooth_scan, list_processes, kill_process, launch_app, open_file, shell_exec

Knowledge / AI: multi_llm_query, knowledge_query, kg_neighbors, kb_info, discover_tools, analyze_path, self_modify

---

Architecture

launch.py               Universal entry point — boots tray, overlay, hotkey, then REPL
install.py              Dependency installer + GPU detection
get-anarchy.ps1         Windows one-liner installer
get-anarchy.sh          Unix one-liner installer
mcp_server.py           MCP server (memory + skills over stdio)

core/
  app_context.py        Dependency injection hub (56 subsystems)
  gateway.py            Non-bypassable audit gateway — single policy source, 7 tiers
  cognition.py          NLP/NLU: intent, entities, SVO, coreference, dialogue state
  nlp_payload_extractor.py  Natural language → tool payload (location aliases, content extraction)
  routing_engine.py     Semantic + lexical tool/skill routing with confidence gating
  execution.py          Tool executor, plan executor, scheduler, sub-agents
  llm.py                Ollama + GGUF backend, model router, cache, token tracker
  memory_facade.py      Unified memory (FAISS vector + SQLite episodic + JSON KV)
  wake_word.py          openwakeword background listener (shared via AudioManager)
  audio_manager.py      Shared mic session — prevents wake word / voice conflicts
  workflow_engine.py    DAG workflow execution with conditional routing + retry
  dev_tools.py          LSP client (Python/TS/Rust/Go/C/Ruby/Lua/PHP), autodoc
  handlers/             REPL command handlers split by domain (11 files)

ui/
  approval_overlay.py   Rich floating approval card — diff viewer, effect + rollback
  tray.py               System tray icon with right-click menu
  hotkey_listener.py    Global Win+Alt+Z wake hotkey

tools/
  [73 tools]            Individual Python modules, one function each
  adapters/
    browser.py          Chrome/Edge via Playwright CDP
    vscode.py           VS Code CLI + UIA adapter
    windows_settings.py PowerShell-backed system settings

skills/                 14 Markdown skill definitions
config/
  permissions.json      Per-tool permission tiers (single policy source)
  config.json           Model, context, temperature, hardware profile
  sandbox.json          Allowed/blocked filesystem paths
  tools_manifest.json   Tool registry (73 tools)
memory/                 FAISS index, SQLite episodic DB, JSON store
models/                 Local GGUF files

---

MCP Server

Exposes ANARCHY's vector memory and skills to any MCP-compatible client (Claude Code, Cursor, ChatOllama):

py mcp_server.py        # Windows
python3 mcp_server.py   # macOS / Linux

Add to Claude Code (~/.claude/mcp.json):

{
  "anarchy": {
    "transport": "stdio",
    "command": "python3",
    "args": ["/path/to/anarchy/mcp_server.py"]
  }
}

---

Models

Default: llama3.2:3b (fast, ~2 GB). Switch anytime:

ollama pull phi4:14b               # strong reasoning, 8 GB VRAM
ollama pull deepseek-coder-v2:16b  # best for code, 8 GB VRAM
ollama pull llama3.1:8b            # balanced, 5 GB VRAM
ollama pull mistral:7b             # solid general purpose
ollama pull phi3:mini              # tiny — works on 4 GB RAM

Set in config/config.json:

{ "ollama_model": "deepseek-coder-v2:16b" }

GGUF fallback: place any .gguf file in models/ and ANARCHY uses it when Ollama is offline.

---

Requirements

• Python 3.10+
• Ollama (auto-installed by the one-liner)
• 8 GB RAM minimum (4 GB workable with phi3:mini)
• Windows 10/11 for native UI features; macOS/Linux for core REPL

Optional language servers for diagnose:

pip install pyright                                      # Python
npm install -g typescript-language-server typescript     # TypeScript/JS
rustup component add rust-analyzer                       # Rust
go install golang.org/x/tools/gopls@latest              # Go
sudo apt install clangd                                  # C/C++
npm install -g intelephense                              # PHP

---

Configuration

config/config.json — model, context length, temperature, GPU layers, threads, hardware profile

config/permissions.json — per-tool permission tiers (the single policy source for the gateway)

config/sandbox.json — allowed and blocked filesystem paths

Secrets stored encrypted in config/secrets.vault:

secret set github_token ghp_yourtoken
secret set HF_TOKEN hf_yourtoken

---

License

MIT — see LICENSE

---

Built for people who want their AI to run on their machine, answer to nobody, and work anywhere.