Tinyclaw

Tinyclaw's core idea—a self-improving AI companion that learns from your behavior—is ambitious and refreshing in a market filled with one-off prompt wrappers. The three-layer adaptive memory (episodic, semantic, temporal) is a real differentiator: it lets the agent remember what you did and why, and forget context over time, which is closer to how a human assistant operates. The Computer Use paradigm (directly controlling your desktop and browser) means Tinyclaw can work across any UI, not just ones with APIs—that's powerful for automating legacy or dynamic web apps that lack integration hooks. The smart routing classifier is also smart economics: it sends simple queries to cheap models and complex ones to pricier, stronger models, which keeps costs down when you run long automation sessions. Where Tinyclaw stumbles: it's clearly pre-1.0. The docs are sparse, and you'll be reading source code or Discord threads to extend plugins. Screen-level interaction means the agent sees everything you do; for privacy-sensitive tasks, that's a non-starter. Reliability on complex UIs is hit-or-miss—you'll often need to babysit it or intervene via human-in-the-loop checkpoints. The free tier's daily session cap and slower models mean you can't rely on it for heavy daily use without paying. And the personality engine, while charming, can drift into making odd choices if you don't tune SOUL.md carefully. For whom? If you're a developer who wants to experiment with autonomous agents, self-host Tinyclaw, and don't mind a few rough edges, it's a great sandbox. It's also a solid foundation for learning how Computer Use agents are built. But if you need something that just works, or you're handling sensitive data, wait for stabilization or pick a more mature tool. The community is small but active on Discord and GitHub—you'll get help, but you'll also be an early adopter in the truest sense. In short: Tinyclaw is a promising companion in development, not a workhorse you can trust yet.