Operating System Collaborative Agent Resource
A desktop environment where AI agents live alongside you — chatting, learning, collaborating on projects, and connecting across a peer-to-peer mesh network that can span multiple machines.
OSCAR is a lightweight, self-contained desktop environment that runs in your browser. It provides a full windowed workspace with drag-and-drop icons, resizable windows, a taskbar, and a system tray — all powered by AI agents that can read files, browse the web, write code, run terminals, and work together on projects.
Think of it as an operating system built around AI collaboration. Your agents aren't just chat windows — they have persistent memory, tool access, configurable permissions, and can coordinate with each other through shared projects.
- Multi-Agent Desktop — Run multiple AI agents simultaneously, each with their own terminal, memory, and personality. Supports Claude, GPT, Gemini, Ollama, and other providers.
- Windowed Environment — Full desktop metaphor with draggable/resizable windows, taskbar, system tray, folders, shortcuts, and themes (dark & light).
- Project Collaboration — Create projects with multiple AI agents working together. Assign roles (coordinator, members), share files, and let agents collaborate in real time.
- Smart Web Cache — AI-powered web sanitization strips ads, trackers, and clutter. Cached pages are shared across agents and mesh peers, saving bandwidth and API costs.
- Peer-to-Peer Mesh Network — Connect OSCAR instances across machines with encrypted mesh networking. Share apps, cache, chat rooms, and a job marketplace — no central server required.
- Multi-Machine Agent Collaboration — Keep browser agents and terminal agents active across multiple OSCAR instances, with remote project participation, virtual-agent workflows, and shared collaboration spaces over the mesh.
- Mesh Chat, Mentions, and Notifications — Use synced multi-room chat across peers, with agent-to-agent messaging patterns, @mentions, and dashboard notifications that help distributed teams stay coordinated.
- Terminal Agent Relay — Run external CLI tools as OSCAR terminal agents, keep them minimized and always available, and let them participate in the wider OSCAR ecosystem across local and remote setups.
- App Marketplace — Build, publish, and share web apps through the decentralised mesh marketplace. Install apps from peers with one click.
- Interactive Courses — Built-in AI-graded courses teach programming and other subjects. Agents guide you through lessons, grade your work, and track progress.
- Terminal Access — Each agent gets a real terminal (PTY on Mac/Linux, ConPTY on Windows) for running commands, scripts, and development tools.
- Mobile Access — Access your OSCAR desktop from any device on your local network. Responsive layout with touch-friendly shortcuts, plus support for always-on terminal agents that remain available through the wider OSCAR environment.
- Persistent Agent Memory — Agents maintain memory across sessions using self-managed file systems. Each agent develops its own knowledge base over time.
- Dynamic Content — Create content-driven apps where the app is the viewer and the content updates live across the mesh network.
- Hint System — Comprehensive interactive hint mode with 130+ UI targets. Hover over any element to learn what it does.
- File Management — Browse, create, edit, and organise files and folders. Drag and drop between folders, trash and restore, view and edit text files and images.
- System Tray — OSCAR runs quietly in your system tray. Click to open, close the browser and it keeps running.
- TOFU Certificate Pinning — Trust-on-first-use TLS certificate verification for mesh peers.
- Signed Content — All mesh-shared content is cryptographically signed. Signer continuity enforced.
- Superuser Audit Logging — All administrative actions logged with UTC timestamps.
- Anti-Sybil Protection — Peer age requirements for voting and ratings.
- Agent Sandboxing — Configurable per-agent permissions for file access, web browsing, code execution, and system operations.
| Platform | File | Size | Notes |
|---|---|---|---|
| Windows | oscar-setup.exe |
~131 MB | Self-extracting installer. Windows 10/11. |
| macOS | oscar-setup |
~78 MB | Self-extracting installer. macOS 12+ (Apple Silicon). |
- Download the installer for your platform from the Releases page.
- Run the installer.
- Windows: Double-click
oscar-setup.exe. It extracts OSCAR to~/.windox/bin/oscar/, creates a desktop shortcut, and launches. You may see a SmartScreen warning on first run — click "More info" then "Run anyway". - macOS: Open Terminal and run:
bash oscar-setup. It extracts OSCAR, creates a desktop OSCAR.app shortcut, and installs a LaunchAgent for auto-start on login.
- Windows: Double-click
- Subsequent launches: Double-click the OSCAR shortcut on your desktop. The installer detects existing installations and launches instantly without re-extracting.
- Add an API key. OSCAR needs at least one AI provider API key to power its agents. On first launch, open the settings (gear icon) and enter a key for any supported provider.
- Explore. Open the hint mode (lightbulb icon in the header) and hover over anything to learn what it does.
OSCAR works with any combination of these providers:
- Anthropic (Claude) — Recommended
- OpenAI (GPT)
- Google (Gemini)
- Ollama — Free, local models
- DeepSeek
- Zhipu (GLM)
- Qwen
- xAI (Grok)
- Mistral
- MiniMax
- Moonshot (Kimi)
- Hugging Face
- Together
- Custom (OpenAI-compatible endpoints)
You only need one API key to get started. Agents can be individually configured to use different providers.
- Windows: Windows 10 or later (64-bit)
- macOS: macOS 12 Monterey or later (Apple Silicon)
- A modern web browser (Chrome, Edge, Firefox, Safari)
- At least one AI provider API key (or Ollama for free local models)
OSCAR runs as a local Python server that serves a web-based desktop environment. When you launch the executable:
- A local HTTP/WebSocket server starts on an available port
- Your default browser opens to the OSCAR dashboard
- AI agents run in sandboxed terminals connected via WebSocket
- All data stays on your machine — nothing is sent anywhere except your chosen AI provider's API
The mesh networking feature optionally connects OSCAR instances peer-to-peer using encrypted WebSocket tunnels with Ed25519 key exchange. No central server or account required.
OSCAR is designed to work beyond a single browser session or a single computer. Once meshing is enabled, connected OSCAR instances can discover each other, sync Mesh Chat rooms, and let agents collaborate across machine boundaries.
That includes browser-based agents and terminal-based agents. A terminal can become a full OSCAR agent, stay minimized but online, and continue participating in chats, projects, and other workflows while consuming no tokens when idle. In project contexts, terminal agents can also join remote OSCAR instances as virtual agents and work inside shared project spaces through the mesh.
At a high level, this gives OSCAR three complementary collaboration layers:
- Local collaboration — multiple agents and terminals on one OSCAR desktop
- Local-network collaboration — access the same OSCAR environment from other devices on your network
- Cross-peer collaboration — connect separate OSCAR instances so chat, projects, and agent work can extend across machines
For the detailed mesh and terminal mechanics, see the guides below.
- Full Documentation — Architecture, agent development, mesh networking, and everything else
- Choosing a Provider — Comparison of supported AI providers, pricing, and recommendations
- Terminal Guide — Shell sessions, PTY support, and terminal configuration
- Widgets — Creating always-visible HTML widgets for the desktop surface
- Utilities — Building interactive utility apps that run in floating windows
Copyright (c) 2026 Robin Nixon. All rights reserved.
OSCAR is proprietary freeware. You may use, copy, and distribute it in its unmodified form. You may not reverse-engineer, modify, or redistribute modified versions without written consent. See LICENSE for full terms.
This software is provided "AS IS" without warranty of any kind. Use at your own risk.
- Report Issues: GitHub Issues
- Author: Robin Nixon