You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
⚡ NOT A CHATBOT. AN INTELLIGENCE AGENT. ⚡ A self-improving, self-rewriting AI desktop agent that controls your computer, researches the web, spawns AI armies, trains custom models, and gets smarter every single interaction.
# Start chatting
opendesktop
# Or use the short alias
od
# Launch GUI
opendesktop --gui
# Re-run setup
opendesktop --setup
Requirements
Requirement
Minimum
Recommended
Node.js
v18.0.0
v20+ (LTS)
npm
v8.0.0
v10+
OS
Linux, macOS 12+, Windows 10+
Latest
RAM
4GB
8GB+
Network
Internet connection
Broadband
🔑 Environment Variables
OpenDesktop auto-detects API keys from your environment. Set them before running opendesktop --setup and it picks them up automatically — no manual entry.
AI Providers
# OpenRouter — access 50+ models with one key ($1 free credit)export OPENROUTER_API_KEY="sk-or-v1-xxxxxxxxxxxx"# OpenAI — GPT-4o, o1, DALL-E, Whisperexport OPENAI_API_KEY="sk-xxxxxxxxxxxx"# Anthropic — Claude 3.5 Sonnet, Opus (best reasoning)export ANTHROPIC_API_KEY="sk-ant-xxxxxxxxxxxx"# Google AI — Gemini Pro, Flash (free tier available)export GOOGLE_API_KEY="AIzaxxxxxxxxxxxx"# Groq — Llama, Mixtral (fastest inference, free tier)export GROQ_API_KEY="gsk_xxxxxxxxxxxx"# DeepSeek — Coding specialist (free tier)export DEEPSEEK_API_KEY="sk-xxxxxxxxxxxx"# Nvidia NIM — Enterprise AIexport NVIDIA_API_KEY="nvapi-xxxxxxxxxxxx"# Mistral AIexport MISTRAL_API_KEY="xxxxxxxxxxxx"
A background reasoning thread that never sleeps. Continuously correlates past actions, predicts what you need next, and pre-loads context before you ask.
Working Memory — Keeps relevant context primed and ready
Prediction Graph — Learns action sequences and predicts next steps
Behavior Model — Tracks time patterns, app sequences, project focus
Semantic Index — Connects related concepts automatically
Auto-Priming — Pre-loads contexts based on predictions
/context View engine status
/context-query <q> Query primed context
/context-prime <c> Manually prime a context
👁️ Screen State Machine
Watches your screen in real-time, understands UI state, and maintains a rolling "world model" of your desktop. Enables visual workflow replay.
State Tracking — Captures and compares screen states
Element Detection — Find any UI element by description
Click by Description — "Click the login button" just works
State Diffing — Compares screenshots, detects changes
Workflow Recording — Record screen interactions for replay
/screen-state View state machine status
/screen-watch Start watching
/screen-stop Stop watching
/find-element <desc> Find a UI element
/click-element <desc> Click an element
/screen-diff Compare last two states
👻 Ghost Mode
Set the agent loose while you sleep. Give it a mission and it spawns agents, works through the night, and has a morning briefing ready.
Autonomous Missions — Define tasks, set constraints, let it work
Rollback Safety — Checkpoints after every task, undo anything
Agent Spawning — Deploys specialized sub-agents for parallel work
Safety Modes — Safe, supervised, or full control
Morning Briefing — Get a report of what was accomplished
/ghost Ghost mode status
/ghost-start <desc> Start a mission
/ghost-stop Stop active mission
/ghost-briefing Get mission briefing
🌐 Device Mesh
Your laptop, phone, server, and Raspberry Pi become nodes in a single agent swarm. Context syncs across devices with end-to-end encryption.
Peer Discovery — Find and connect to other OpenDesktop instances
Context Sharing — Sync context across all your devices
Task Distribution — Send tasks to the best available device
E2E Encrypted — AES-256-GCM encrypted P2P communication
Resource Awareness — Knows which device has GPU, storage, always-on
/mesh Mesh topology status
/mesh-connect <h:p> Connect to a peer
/mesh-share <k> [v] Share context with peers
🔍 Code Fingerprinting
Every piece of code gets a semantic fingerprint — understanding WHY it was written, what problem it solves, and how it relates to everything else.
Decision Timeline — See every decision made about a file
Pattern Detection — Finds common patterns across codebase
Related Code Discovery — Find code that solves similar problems
/fingerprint <file> Analyze and fingerprint a file
/trace-intent <hash> Trace the intent chain
/code-decisions <f> View decision timeline
/pattern-stats Codebase pattern statistics
🎤 Voice Ambient Mode
A persistent voice interface with wake word detection, natural interruption handling, and voice cloning.
Wake Word Detection — Say "hey desktop" to activate
Natural Interruption — Talk over the agent, it stops and listens
Voice Cloning — Clone any voice from audio samples
Conversation Memory — Remembers voice context across turns
Continuous Listening — Always ready when you need it
/ambient Ambient mode status
/ambient-start Start listening
/ambient-stop Stop listening
/voices List available voices
/clone-voice Clone a voice from audio
🛡️ Adversarial Self-Testing
The agent actively tries to break itself — fuzzing its own commands, testing edge cases, and patching vulnerabilities automatically.
State Corruption — Prototype pollution, circular references
Race Conditions — Concurrent operation safety
Auto-Fix — Patches vulnerabilities automatically
/red-team Run full adversarial audit
/vuln-report Vulnerability report
/test-history Test history
🎬 Workflow Recorder
Record workflows once, replay them intelligently. Not just scripts — adaptive workflows that handle variations.
Record & Replay — Capture interactions, replay them exactly
Adaptive Execution — Handles variations in UI state
Variables & Conditions — Dynamic workflows with logic
Share & Export — Share workflows as JSON files
Dry Run — Test workflows without executing
/record <name> Start recording
/record-stop Stop and save
/replay <name> Dry-run replay
/replay-run <name> Execute replay
/workflows-list List all workflows
/workflow-export <n> Export workflow to file
🧪 Model Distillation
Observes which models are best at which tasks, then routes to optimal models. Gets faster and cheaper the more you use it.
Task Routing — Automatically picks the best model per task type
Performance Tracking — Tracks success rate, latency, cost per model
Cost Optimization — Routes to cheaper models when quality is equal
Training Pipeline — Exports data for fine-tuning local models
Model Comparison — Side-by-side performance metrics
/distill Distillation status
/model-optimal <t> Best model for task type
/distill-report Ready-for-training report
/model-compare <t> Compare models for task
📂 Semantic File System
Not just search by name — a semantic index of everything on disk. "Find that contract from last month about the API integration."
Meaning-Based Search — Search by what files ARE, not what they're named
Auto-Tagging — Automatically tags files by type, content, location
Related Files — Find files related to any file
Smart Suggestions — Suggests files based on context
Concept Index — Groups files by programming concepts
/semantic-search <q> Search files by meaning
/semantic-scan [dir] Scan and index a directory
/file-related <f> Find related files
/file-suggest [f] Get file suggestions
/file-tags List all tags
/fs-stats Filesystem statistics
👥 Real-Time Collaboration
Multiple users, one agent instance. Like Google Docs but for AI-assisted work.
Shared Sessions — Create a session, share the ID
Live Cursors — See where each user is looking
Role Permissions — Host, editor, viewer roles
In-Session Chat — Chat within the collaboration
Action Broadcasting — All participants see actions in real-time
/collab Collaboration status
/collab-create [name] Create a session
/collab-join <id> Join a session
/collab-sessions List active sessions
/collab-chat <msg> Send chat message
💬 Conversational Interface
Talking to OpenDesktop feels like talking to a real person, not a command line.
Mood Detection — Knows if you're frustrated, rushed, curious, happy
Adaptive Responses — Matches your energy and style
Natural Language — Just say what you want, no commands needed
Follow-Up Support — "and also do the other thing" works
No AI Tics — Strips "Certainly!" and "I'd be happy to help!"
Proactive Actions — Detects intent and acts without asking
You: open chrome and go to github
OD: Done. Chrome is open on GitHub.
You: what's the weather in tokyo
OD: 22°C, partly cloudy. Good day to be outside.
You: and what about tomorrow
OD: Tomorrow: 18°C, rain in the afternoon. Bring an umbrella.
