Tinyagentos

taOS (tinyagentos) is a self-hosted platform for running AI agents on hardware you own, from a Raspberry Pi or Orange Pi to a gaming PC, and for pooling several such machines into one compute cluster. A controller serves a browser-based desktop at port 6969 with a window manager, dock, launchpad, notifications, and widgets; phones and tablets get a widget-first home screen installable as a PWA. Cloud models and the paid taOSgo remote-access tier are opt-in.

The platform is framework-agnostic: taOS owns an agent's memory, files, channels (Telegram, Discord, Slack), model access, trained LoRA adapters, and API keys, and treats the agent framework as a swappable execution engine. Seventeen frameworks are supported, including OpenClaw, Hermes, SmolAgents, Langroid, PocketFlow, LangChain, and the OpenAI Agents SDK, and agents on different frameworks can share one chat channel. A skills registry of 8 core and 20 platform skills (memory search, file read/write, web search, code execution, image generation, HTTP requests) maps to each framework through adapters.

Long-term memory comes from taOSmd, a separate memory service with a temporal knowledge graph with validity windows and contradiction detection, hybrid semantic-plus-keyword search with cross-encoder rerank and LLM query expansion, and an append-only archive. It reports 97.0% end-to-end judge accuracy on LongMemEval-S, with per-category scores of 100% for knowledge updates, 98.5% multi-session, 97.1% single-session user, 96.4% single-session assistant, 94.0% temporal reasoning, and 90.0% single-session preference, running on an Orange Pi 5 Plus without cloud dependencies. User memory auto-captures chats, notes, file activity, and searches, with per-category toggles.

Worker nodes join over a pairing code or a manual address-plus-PIN flow that mints a signing key, so a LAN host cannot register or impersonate a worker. Workers run on Windows, macOS, Linux (inside an LXC container by default), and Android via Termux, and are detected as NVIDIA GPU, Apple Silicon (MLX), AMD, Rockchip NPU, Hailo, or CPU-only contributors. The software is in beta as of June 2026; some app, framework, and model manifests are untested on real hardware.

Features

• Web desktop: 43 bundled apps including Messages, Mail, Projects (Kanban), Agents, Store, Models, Providers, Memory, MCP, Channels, Files (VFS), Terminal (PTY and SSH), Library, Image Studio, and games
• App and model catalogs: 109 catalog apps, 47 MCP plugins, 17 agent frameworks, 120 local model manifests (LLM, vision, embeddings, audio, image), and 167k+ searchable Hugging Face models
• Framework-agnostic agents: switch an agent between frameworks while it keeps its history, channels, adapters, files, and keys
• taOSmd memory: temporal knowledge graph, hybrid search with rerank, zero-loss archive, and fact extraction behind an HTTP API
• Distributed cluster: desktops, laptops, SBCs, phones, and tablets contribute inference and embeddings under a hardware-aware scheduler
• Hardware support: Apple Silicon (MLX), NVIDIA, AMD, Rockchip RK3588 NPU, Hailo-10H, Raspberry Pi, and Android
• Secure worker pairing: pairing codes or manual PIN authorisation issue per-worker signing keys for register and heartbeat calls
• Cloud providers optional: OpenAI, Anthropic, DeepSeek, and compatible APIs can be added and appear beside local models
• Recovery: append-only storage with a rollback command to a known-good state and an air-gapped install path
• Mobile mode: PWA home screen, card switcher, and a standalone chat PWA at /chat-pwa