Tinyagentos

taOS is a self-hosted platform for running AI agents on hardware you already own: an old laptop, a Raspberry Pi, an Orange Pi, a gaming PC, an Apple Silicon Mac, or several of them joined into one compute cluster. Agent memory, conversations, files, and model access stay on your machines, and cloud models or paid remote access are opt-in rather than required.

The platform is framework-agnostic. taOS owns the parts of an agent that persist (memory, files, messaging channels, secrets, model access, and configuration) and treats the agent framework as a replaceable execution engine that runs inside an isolated LXC or Docker container. Switching an agent from Smolagents to LangChain to OpenClaw keeps its history, its Telegram, Discord, and Slack connections, its trained LoRA adapters, and its API keys. Seventeen frameworks are installable from the store, including OpenClaw, Hermes, PicoClaw, DeerFlow, Langroid, PocketFlow, Agent Zero, Moltis, and the OpenAI Agents SDK, and six of them have been verified in a shared multi-framework group chat.

Everything is driven from a browser-based desktop: a window manager, dock, launchpad, widgets, notifications, and 43 bundled apps, which swap to a widget-first home screen on phones and tablets and install as a PWA. Memory is provided by taOSmd, a temporal knowledge graph combined with hybrid semantic and keyword search, a cross-encoder reranker, and a zero-loss append-only archive, which reports 97.0% end-to-end judge accuracy on LongMemEval-S while running on an Orange Pi 5 Plus.

Inference is unified behind a LiteLLM proxy that exposes one OpenAI-compatible API with a virtual key, budget, and rate limit per agent, so swapping a local Ollama backend for a cloud provider changes no agent configuration. A curated catalog of 120 model manifests covers LLMs (Qwen, Llama, Gemma, Phi, Mistral, DeepSeek, Granite, Command-R), vision, embedding, reranker, speech, and image-generation models, with Rockchip RK3588 NPU and Hailo-10H variants, and a live browser searches Hugging Face and the Ollama library. Exo can be added to split a single large model across several devices, and TurboQuant KV-cache quantization extends context length on consumer GPUs.

Supported hardware spans Apple Silicon (Ollama or MLX), NVIDIA (CUDA or Vulkan), AMD (ROCm), Rockchip NPU boards, Raspberry Pi with or without accelerator HATs, Android phones as workers, and CPU-only devices, mixed freely in one cluster. The project is in beta, and some app, framework, and model install manifests are untested on real hardware.