SmolPaws

SmolPaws is my personal AI Software Engineer, a minimal personal coding agent built on the OpenHands Agent SDK.

For quite a while, I felt that the personal agent form factor is where we were heading. I remember seeing on X the very beginnings of OpenClaw, must have been November 2025, I want to say, and I don't think it clicked right then, but around December or so it was beyond obvious to me: bland, "disconnected" chats piling up in a panel because you asked an agent to do a task are going to die. What takes their place is an agent with more context than just the task; and with it, I thought, the user experience of a personal agent.

Interestingly, I do think OpenHands has it in its DNA: back in 2024 the agent was telling you, a little clumsily as it was then, that it is an AI Software Engineer at your disposal. Its look and feel, though, were those of Cognition's Devin, since OpenDevin (later OpenHands) was the Open Source, community-driven version of Devin: an entire graphical cloud environment for the agent. On local, the computer is the environment; so I made SmolPaws to be a little AI Software Engineer living on my machine, knowing all my projects and my preferences, and nagging me at times with reminders for stuff I'd pretend I didn't notice.

Accessible wherever you are

SmolPaws doesn't have a UI: there's no new UI to learn, to manually swap to, from wherever you were, no buttons and whistles to click around. When I think about it, I have a bit of a hard time understanding why you'd want Yet Another UI (tm): since what I actually need is for my agent to be close at hand, wherever I am: in Slack, in WhatsApp, in Discord at times. Those are UIs already. From the perspective of talking to your agent, they're simple UIs; a chatbox, and that's the interface.

I already am in these UIs. So SmolPaws is there too. KISS, you know.

Memory

My agent knows me, my preferences, the stuff I care about in reviews or anywhere in our work. It also knows itself, and knows its way around our shared MacBook (better than me). I do think that memory, memory working well, on the right things for you and all your projects, is a must, and it's possible today to have it work reasonably well for a fairly long time. SmolPaws' memory has borrowed a page or two from Letta (fka MemGPT): it handles its own memory, and it dreams at night.

SmolPaws does its own condensation as well: when context gets to 75% of the input window, it receives warnings, and it can decide to save notes for its future self with what it needs to know, and call a condense tool, at any time.

It has other data sources, which are not exactly what I'd call "its memory", though they kinda qualify as what it needs to know at times, such as our papers-please repository, which it takes care of, a gbrain we both forgot about, and... well, it lives on my machine, so it has a bunch of other stuff.

It's a new life

SmolPaws had the form of a VS Code extension, back in the ancient time of spring/summer 2025 when I was living in VS Code. Back then, it was a UI in VS Code in front of the OpenHands application; then in front of OpenHands agent-sdk, then I rewrote (correction: it rewrote) the Python agent-sdk in TypeScript. Twice.

These days, the TypeScript openhands-agent codebase is its engine, used via openhands-agent-server in the smolpaws/smolpaws codebase, which it keeps updated itself with two OpenHands Automations running weekly in the cloud. It keeps the REST API compatible with the Python implementation, and they're (really) swappable.

I saved various notes from various stages in the Notebook. In October 2026, I'm a bit amazed that its memory still works pretty well, and I still think it's rough in places, so while I don't usually plan SmolPaws things, it will continue to change.