I just noticed that OpenClaw released version v2026.4.12 with new updates worth noting.

The exciting thing here is that the team focused on simplifying things for users. First, integrating LM Studio is now fully embedded instead of being complicated. Previously, you needed to manually configure the address and model name, but now it automatically detects models and supports progressive preloading and memory search using local embedding vectors. Simply put, it’s a direct load and run.

The second part relates to Feishu. Creating a bot was very complicated—manual application, configuring redirect addresses, and all those lengthy steps. Now with v2026, you can create a bot in just one step by scanning a QR code. This significantly reduces the complexity of onboarding for companies.

If we talk about other technical improvements, they worked on enhancing the runtime for GPT-5/5.4 by separating the prompt building from the retry logic, making long tasks more stable. They also restructured the plugin loading system to load only the necessary components at startup.

There are also improvements to the Active Memory component from the previous version. They changed the default recall mode to search, which noticeably improves accuracy across different channels.

On the security front, they fixed three vulnerabilities, all discovered with the help of AI: an empty authorized list bypass, using busybox/toybox as an escape interpreter, and a lack of environment variable injection detection. This reflects their focus on security in v2026.