jaylfc/taOS

What it is: taOS is a self-hosted AI agent platform that turns your own hardware (old laptops, Raspberry Pis, gaming PCs, even Android phones) into a private, distributed AI compute cluster. It provides a full web-based desktop environment to manage agents, models, memory, and files—all running on machines you own, with cloud services strictly optional.

Core idea: Data sovereignty. Your agents' long-term memory, conversations, and files live on your hardware (a £170 Orange Pi is cited as sufficient for the memory system), not on someone else's servers. It's open source (AGPL-3.0) and supports air-gapped installs.

Key features (all from the README):

• Web Desktop OS: A browser-based environment with a window manager, dock, launchpad, and 43 bundled apps (chat, mail, file manager, terminal, image studio, etc.). It adapts to a widget-first mobile/tablet layout and is installable as a PWA.
• Distributed Compute Cluster: Connect diverse devices as workers (Windows, macOS, Linux, Android via Termux). Workers pair securely via a one-time code to prevent unauthorized access. It detects NVIDIA GPUs, Rockchip NPUs, and falls back to CPU-only workers.
• Framework-Agnostic Agents: You can switch between 17 agent frameworks (e.g., SmolAgents, LangChain, OpenClaw) without losing the agent's memory, files, API keys, or channel connections (Telegram/Discord/Slack). The platform manages the lifecycle outside the framework.
• taOSmd Memory System: A self-hosted, framework-agnostic memory system using a temporal knowledge graph, hybrid vector search, and an append-only archive. The README claims 97.0% end-to-end accuracy on the LongMemEval-S benchmark, running entirely on local hardware.
• Large Catalogs: Includes 109 catalog apps, 47 MCP plugins, a curated local model catalog (120 manifests for LLMs, vision, embeddings, audio, image gen, including NPU variants), plus search over 167k+ HuggingFace models.
• Skills & Plugins: A framework-agnostic skill registry (e.g., memory_search, file_read, web_search) that translates across all supported frameworks.

Current status: Beta software (as of 2026-06-02). The core backend, API, memory system, and multi-framework chat work, but the desktop GUI has some rough edges. Many install manifests for the 100+ apps and models haven't been tested on real hardware yet, so some may fail to install—the maintainer asks users to report failures via GitHub issues.

Target user: Tech-savvy individuals who want full control and privacy over their AI stack, are comfortable running install scripts on Linux/macOS/Windows, and are willing to tinker with beta software.