I followed the launch of OpenClaw v2026 and found the changes they brought this time quite interesting. People working with local models should be enjoying this integration with LM Studio, because before it was that manual work of configuring the model address and name. Now, automatic model discovery comes ready out of the box, with vector-based memory search and everything else.

What drew the most attention was the simplification in Feishu. In the past it was a tedious process: you’d go to Plataforma Aberta, create the app manually, configure the callback... now it’s just scan a QR code and you’re done—your bot is up and running. They significantly lowered the entry barrier for people working on enterprise integrations.

There are a few more great things in v2026.4.12 that deserve highlighting. They improved the execution time of GPT-5/5.4 by separating prompt construction from the retry logic, so longer tasks became more stable. And that Active Memory plugin that came before? They enhanced the search—now it’s in search mode by default, which increases accuracy across different channels.

They also paid close attention to plugin loading. Now it only activates the components that are declared in the manifest on initialization, so it no longer loads all that unnecessary stuff. Oh, and they fixed three security vulnerabilities that the AI helped identify—such as that story of an empty list of approvers bypassing verification, using busybox as the interpreter, and injecting environment variables.

If you work with open-source AI agents, it’s worth taking a look at what changed in this version.