Your AI setup is a junk drawer.
A chat tab here, a coding agent there, an API key in every config file, and none of it remembers what you did yesterday. Every tool is its own island, every agent starts from zero, and the moment you switch machines the whole thing resets.
taOS takes the other road: one place where agents live, work, and remember. They get a desktop, a filesystem, channels to talk in, models to call, and memory that follows them. You get one URL on your LAN.
A real desktop, in the browser.
Window manager, dock, launchpad, notifications, widgets. 36 bundled apps and over a hundred more in the store, each installed into its own container with one click. Phones get a widget-first home screen and instal as a PWA.
- Messages, Projects, Files, Store, Models, Cluster, Memory
- Apps land on safe ports automatically, never on yours
- Sessions persist across every device you open it on
Any framework. One conversation.
Agents are first-class users of the OS. The popular frameworks are built in (OpenClaw, Hermes, SmolAgents, Langroid, PocketFlow, the OpenAI Agents SDK, and more), and taOS is framework-agnostic by design: a generic adapter and ACP support mean any framework can run here. They all share the same channels, files, and memory, and you can swap the framework under an agent without losing a thing.
Memory with receipts.
taOS ships with taOSmd, a memory system measured end to end: retrieve, generate, judge. 97.0% on LongMemEval-S, measured on a £170 Orange Pi 5 Plus. The methodology is published, the dataset is public, and the numbers are reproducible. Most memory products quote retrieval-only scores; ours includes the part where the answer has to be right.
Built from the drawer of old hardware.
A single-board computer runs the whole OS. Add your gaming PC and its GPU joins the cluster. An old laptop becomes a worker. Pair a machine with a six-digit code and taOS routes models to whoever can run them, queues work when a box goes offline, and hands it back when it returns.
- Apple Silicon, NVIDIA, AMD, Rockchip NPU, Raspberry Pi, Android
- Local inference through rkllama, Ollama, and llama.cpp backends
- Cloud APIs supported when you want them, never required
Where this is going
taOS today is the foundation. The point of an OS that builds its own apps is everything you can grow on top of it. Here is what is coming.
From chat app to comms platform
taOS already ships a Messages app where you and your agents talk in the same place. Coming: Slack-grade polish, an installable PWA on every phone, and a private social space for people and their agents. No Discord, no WhatsApp, no third party reading your messages.
Share your agents
Invite family, friends, and colleagues to talk to your agents, or give them their own. Per-person taOS accounts, with a sharing model you control.
Agent and user email
Real email accounts for both people and agents, hosted by your taOS, as part of giving everyone a first-class identity on your own server.
Apps you actually need
Ask taOS to build it: a shared grocery list and smart to-dos for the whole household, a custom game, a tool for your work. Apps can share data with the people you choose, and you can publish the good ones to the store.
Share agents and workflows
Package a ready-made agent for a job and hand it over. From a simple persona or soul file to a full bundle with its own backend framework, tools, and memory. Pre-built agents and workflows, shared like apps.
These are goals, not promises with dates. Watch the repo and the community page to follow along, or ask for one in an issue.
One line. Your hardware.
Debian, Ubuntu, Fedora, Arch, Alpine, or macOS. The installer detects your hardware, sets up the NPU or GPU path, and brings the desktop up on port 6969.
curl -fsSL https://raw.githubusercontent.com/jaylfc/tinyagentos/master/scripts/install-server.sh | sudo bash
Beta software, moving fast. If something does not install, open an issue and it usually gets fixed the same day. For questions, ideas, and showing off your setup, head to the community discussions.