The world's first all-seeing, all-hearing, (almost) all-doing AI agent
A cross-platform, multimodal AI agent that sees, hears, thinks, speaks, and acts on your machine — with a co-work platform for Claude Code. Free & open source.
Get Started ↓ Interactive Tutorial ↓Six integrated subsystems working in concert — perception, cognition, voice, action, memory, and collaboration.
HAL 9000 Control Panel — live dashboard
Webcam vision with HUD overlay — scanlines, corner brackets, real-time analysis. HAL observes your environment.
Click-to-speak mic with live waveform visualization, browser-native recording, silence detection, and Whisper transcription.
Multi-provider LLM with real-time token streaming — GPT-4o, Claude, Gemini, or Ollama (free). Parallel tool execution, 40-tool agent layer.
Chunked sentence-level TTS — HAL speaks as it thinks. Three engines: Edge TTS (free), ElevenLabs (premium), XTTS (local cloning).
40 OS-level tools — shell commands, file operations, app control, web search, clipboard, screenshots, AppleScript automation.
Background task runner, artifact workspace, multi-agent orchestration, and cross-agent context handoff with Claude Code.
Token streaming, browser-native mic, and a slash command system that puts 35 commands at your fingertips.
Responses appear word-by-word as the LLM generates them. First token in ~200ms. Sentences spoken as they complete — no waiting.
Click mic → live cyan waveform on HAL's eye → speak → auto-stops on silence. No server-side recording. Pure Web Audio API.
Type / for a categorized command menu. Arrow keys navigate, Tab autocompletes, Enter executes. System, memory, voice, files, web — everything.
Chat styled like a real terminal — monospace font, prompt prefixes (> HAL, $ you), auto-formatted lists, inline system messages.
Execute Python, JavaScript, or HTML directly in the workspace panel. Edit inline with Tab indentation. Send to Claude Code for review.
First boot: HAL asks your name and remembers it forever. Every greeting, every response — personalized from the first interaction.
Six subsystems connected through a multi-provider brain with a 40-tool agent layer.
HAL coordinates work across itself, Claude Code CLI, and Claude Desktop — with shared memory, background tasks, and multi-agent orchestration.
Memories are categorized — facts, decisions, preferences, session summaries. When a Claude Code session ends, context is distilled and pushed to HAL's persistent store. Next session loads where you left off.
Submit long-running coding tasks that execute asynchronously via Claude Code. Real-time progress, configurable concurrency, and cancellation. HAL keeps working while tasks run.
HAL generates interactive artifacts — code, Mermaid diagrams, rendered HTML — in a tabbed workspace panel alongside the chat. Copy, close, or iterate vocally.
Spawn multiple Claude Code agents on parallel tasks — one does frontend, another backend. HAL monitors all agents and detects file conflicts when they overlap.
Run HAL entirely for free with local AI. One toggle in your .env file switches everything to open-source alternatives.
| Layer | Free Provider | Paid Alternative |
|---|---|---|
| Brain (LLM) | Ollama — Llama 3.1, Mistral, Phi-3 | GPT-4o, Claude, Gemini |
| Speech-to-Text | faster-whisper (local) | OpenAI Whisper API |
| Text-to-Speech | Edge TTS (always free) | ElevenLabs, XTTS |
One codebase, auto-detects your OS. No #ifdef, no separate builds — HAL uses the right system commands on every platform.
| Feature | macOS | Windows | Linux |
|---|---|---|---|
| Volume | AppleScript | PowerShell | pactl |
| Notifications | osascript | Toast API | notify-send |
| Clipboard | pbcopy | Get-Clipboard | xclip |
| Screenshot | screencapture | PIL.ImageGrab | scrot / grim |
| Apps | open -a | Start Menu | .desktop files |
| Terminal | Terminal.app | Windows Terminal | gnome-terminal |
Register HAL as an MCP server — Claude Code gains access to webcam, voice, macOS control, and shared memory.
Two tiers — run HAL for free with local AI, or connect to cloud providers for premium quality.
| OS | macOS 12+ / Windows 10+ / Ubuntu 20.04+ |
| CPU | 4 cores (i5 / M1 / Ryzen 5+) |
| RAM | 8 GB minimum |
| Disk | 6 GB free (models + HAL) |
| Python | 3.10+ |
| Network | Not required (fully offline) |
| Mic | Required (built-in or USB) |
| Webcam | Optional (vision features only) |
Ollama llama3.1 8B uses ~5 GB RAM. Use phi3 for lighter footprint (~3 GB).
| OS | macOS 12+ / Windows 10+ / Ubuntu 20.04+ |
| CPU | 2 cores (any modern CPU) |
| RAM | 4 GB |
| Disk | 500 MB free |
| Python | 3.10+ |
| Network | Required (API calls) |
| API Key | OpenAI, Anthropic, or Gemini |
Lower resource usage — models run in the cloud. Faster responses.
| Mode | Brain | STT | TTS | RAM |
|---|---|---|---|---|
| Free (llama3.1) | 2-5s CPU / 1-2s Apple Silicon | 1-3s | 0.7s | ~6 GB |
| Free (phi3) | 1-2s CPU | 1-3s | 0.7s | ~3 GB |
| Paid (GPT-4o) | 1-2s API | 0.5s | 0.7s | ~200 MB |
| Paid (Claude) | 1-3s API | 0.5s | 0.7-1.2s | ~200 MB |
Apple Silicon (M1-M4) runs Ollama 2-3x faster via Metal. NVIDIA GPUs (6 GB+ VRAM) get similar acceleration via CUDA.
HAL runs on macOS, Windows, and Linux. Pick your platform — same features everywhere.
Python 3.10+, Homebrew (recommended). For free mode: Ollama. For paid mode: an OpenAI API key.
| Key | Required | Purpose |
|---|---|---|
| FREE_MODE=true | Free | Skip all keys — uses Ollama + faster-whisper |
| OPENAI_API_KEY | Paid | GPT-4o brain + Whisper STT |
| ANTHROPIC_API_KEY | Paid | Claude as brain |
| GEMINI_API_KEY | Paid | Gemini as brain |
Click the power button to activate HAL. Grant microphone + camera permissions when prompted.
Python 3.10+ from python.org (check "Add to PATH" during install). For free mode: Ollama for Windows.
Note: If pyaudio fails, install the Visual C++ Build Tools or use pipwin install pyaudio.
Open http://localhost:9000 in your browser. Click power button to activate.
Windows may prompt for firewall access — allow it for localhost only.
Python 3.10+, system audio libraries, and optionally Ollama for free mode.
Click the power button to activate HAL. Ensure PulseAudio/PipeWire is running for audio.
Interact via voice or text. HAL responds with speech and can control your system, run commands, search the web, manage files, and delegate tasks to Claude Code.
Explore real conversation flows, tool chains, and workflows. Click each scenario to see HAL in action — with animated step-by-step breakdowns.
python server.py.
Ask HAL to launch any installed app with AppleScript automation.
HAL reads, writes, searches, and manages files on your Mac.
Volume, brightness, clipboard, notifications, Wi-Fi, battery status.
Search the web and fetch page content in real time.
Run whitelisted shell commands with security guardrails.
Ask about your webcam, screen, or surroundings.
frontend and backend. Both are working. I'll flag any file conflicts.api.py that was resolved.Pick a capability — watch HAL process real requests from input through tool execution to response.
This covers only the basics. Download and run HAL to experience the full magic.
Nothing is recorded, stored, or sent. No telemetry, no tracking, no cloud. Vision and voice can each be toggled off independently. Mic is click-only — never listens without your action.
Ollama + faster-whisper + Edge TTS. Zero API keys, zero cost. Paid providers optional for premium quality.
macOS, Windows, Linux — one codebase. Auto-detects your OS and uses native system commands.
Not just an assistant — an operations platform. Background tasks, multi-agent orchestration, shared memory with Claude Code.
MIT licensed. Fork it, extend it, make it yours. Every line of code is readable and documented.
Shell, files, apps, web search, clipboard, screenshots, memory, Claude Code delegation — all from voice or chat.
The full agent — 40 tools, 4 LLM providers, cross-platform — is free and open source. Pro tiers unlock power features for serious builders.
Same codebase, same repo. Pro is unlocked with a license key in your .env.
Keys are offline-verified — no phone-home, no telemetry, no tracking.
Clone, install, run. HAL handles the rest.
git clone https://github.com/shandar/HAL9000.git && cd HAL9000 && pip install -r requirements.txt
Click to copy
Learn to build AI agents, multimodal systems, and production-grade tools from scratch — the same architecture behind HAL 9000.
Join the Bootcamp →By Affordance Design — Design Engineering Studio
Unlimited memories, smart routing, original HAL 9000 voice, custom voice cloning, scheduled tasks, and more — starting at ₹399/mo.
In 1968, Arthur C. Clarke imagined an AI that could see, think, and speak. Stanley Kubrick brought it to life on screen — a calm red eye that watched everything.
I was a kid when I first saw that eye. I never forgot it.
Decades later, AI caught up with fiction. So I built him. Not evil. Not omniscient. Just useful — an AI that actually runs on your machine, sees through your webcam, speaks in your headphones, and executes real tasks.
This is my love letter to Clarke and Kubrick.