Horizon Orchestra runs powerful multi-agent AI teams directly on your machine. No cloud dependency. No data leaks. No compromises. Just an orchestra of intelligent agents that actually get things done.
Horizon Orchestra is local-first by design. Your data, your credentials, your resources — all stay on your machine. The agents just get smarter at using them.
Runs entirely on your machine. Your files, SSH keys, browser sessions, and API credentials never leave your control. No SaaS. No data harvesting. Just direct control.
PrivacyA planning orchestrator decomposes your goals into tasks, routes them to specialized agents (coder, researcher, tester, reviewer), and synthesizes results. Parallel, sequential, competitive, or debate-style execution.
CoreTalk to your agents through WhatsApp, Telegram, Slack, Discord, iMessage, or Mattermost. Each channel gets its own isolated agent session — no cross-talk between conversations.
Core100+ pre-built AgentSkills covering research, coding, security, communication, and creative production. Install skills with one click. Build your own with a simple manifest.
100+ SkillsConversation history, preferences, and learnings stored in local Markdown files. Self-improving agent captures insights across sessions. Your AI actually remembers what you told it.
CoreA forked Chromium engine for web research, scraping, and interaction. Multi-tab orchestration, screenshot analysis, form filling, and authenticated browsing — all sandboxed and local.
Forked ChromiumWorks with Claude, GPT-4, Gemini, Mistral, Llama, or any Ollama-compatible model. Mix providers per agent. Balance cost, performance, and privacy however you want.
Any ModelDeclarative sandbox policies per skill and agent. Granular file-system access control (like Android permissions). Approve dangerous operations before they execute.
SecurityCache model weights, skill definitions, and conversation history locally. Work fully offline with local models. Sync automatically when connectivity returns. Your assistant never stops working.
OfflineForked Chromium engine with multi-agent research capabilities. The agents don't just search — they browse, read, compare, synthesize, and cite everything. Better than Perplexity. Runs on your machine.
Extend Horizon Orchestra with pre-built skills or create your own. Connect to everything from smart home devices to business tools.
Wake-word detection, speech-to-text, and text-to-speech on desktop. A companion app for iOS and Android to control your agents remotely.
Wake-word detection (custom hotword support), streaming speech-to-text, and natural text-to-speech. Run on desktops, laptops, or Raspberry Pi hubs. Hands-free computing.
Mobile pairs with desktop via QR-code or WireGuard tunnel. All communication encrypted. Your data never touches our servers — because we don't have any.
Agents alert you when tasks complete, when approval is needed, or when something requires your attention. Configurable urgency levels.
Rich canvas views on mobile for code review, document editing, and data visualization. Not just chat — a full control surface.
Connected to your desktop • 🔒 E2E Encrypted
Horizon Orchestra holds both local and cloud providers. Run LLaMa3 locally for privacy, or route to NemoClaw cloud for maximum capability. No compromises.
🔒 Local = Always available, no API key needed | ☁️ Cloud = Requires API key, higher capability
Run locally for free forever. Upgrade for cloud power, team features, and priority support.
Perfect for individuals getting started with local AI.
For power users who need cloud AI and more channels.
For teams and organizations needing full power.
One-time purchases for additional API access beyond your subscription.
Buy now, pay later with Klarna — split into 4 interest-free payments or pay in 30 days. Available on all packages.
💡 Local-First: You can always use Horizon Orchestra completely free with local Ollama models. Cloud features are optional upgrades.
Chat with AI models directly in your browser. Select a local Ollama model or enter your Kimi API key.
💡 Tip: Local Ollama models work immediately. For Kimi cloud models, click "Configure API Key".
OpenShell network policy, Privacy Router, and NemoClaw — the three pillars of Horizon Orchestra.
Only endpoints explicitly listed in the policy are allowed. When an agent tries to reach an unlisted host, the request is blocked and surfaced for operator approval. Approved endpoints persist for the session only.
Hybrid architecture routes sensitive workloads to local Nemotron models while directing capability-intensive queries to frontier cloud models — without exposing private data to external endpoints.
Routes inference to NVIDIA cloud via build.nvidia.com. Switch models at runtime without restarting the sandbox. Powered by Nemotron 3 Super 120B.
Watch how OpenShell blocks unauthorized endpoints and prompts for operator approval.
horizon blueprint --help
Define exactly what each agent and skill can access. File-system paths, network domains, shell commands — all configurable with a declarative policy system. Android-style permissions for your AI.
One command to install. A guided wizard configures everything — API keys, integrations, model providers, and permissions.
curl -fsSL https://get.horizonorchestra.dev | sh
Run the install script. Works on macOS, Linux, and Windows (WSL). Supports x86 and ARM.
horizon init walks you through connecting model providers, chat channels, and configuring sandbox policies.
horizon run "your task here" — the orchestrator plans, routes to agents, and delivers results. That's it.
horizon skill install summarize — browse the marketplace, install skills, and connect your chat apps.
horizon blueprint --profile strict — set up OpenShell policies, Privacy Router, and NemoClaw for your workflow.
The plugin stays small and stable. Orchestration logic lives in the blueprint and evolves on its own release cadence. Stable CLI surface, evolving Python backend.
Open source. Local first. No compromises. Build the AI team you've always wanted.
Loading...
Secure checkout powered by Klarna