Introduction

The backbone and soul of my personal assistant, OpenClaw, has evolved once again! It feels like I just applied the 2.22 update yesterday, but 2.23 is already here.

Rather than swapping out the brain entirely, this 2.23 update is the ultimate refinement in details and stability—it focuses on “saving unnecessary API costs, keeping the assistant’s mouth shut when it should be, and locking down the system flawlessly against external threats.”

As always, from the perspective of a “non-CS major / vibecoder”, let’s quickly go over the core features you can feel right away. Let’s go! 🏃‍♂️💨

1. 🧠 New Toys: Kilocode & Moonshot(Kimi) Support

Great news for those who love swapping AI models in OpenClaw.

  • Official Kilocode Support: You can now plug in powerful models like kilocode/anthropic/claude-opus-4.6 as your default. Authentication and setup run buttery smooth.
  • Enhanced Moonshot(Kimi): China’s powerful Moonshot model now has native support for its web search pipeline and video analysis features. Throwing a video at Kimi for analysis is now much more natural.
  • Vercel AI Gateway: When routing Claude models through Vercel, the system now smartly recognizes shorthand names and normalizes them perfectly.

2. 💸 Protecting Your Wallet: Prompt Caching & Memory Optimization

If you’ve ever been hit by an API cost bomb, this update is pure honey.

  • Precision Prompt Caching Control: You can now configure cache retention time (cacheRetention) separately per model and per agent. No more wasting expensive cache on trivial chatter! (Moonshot and ZAI/GLM now benefit from caching too.)
  • Optimized BOOTSTRAP Loading: Previously, modifying AGENTS.md or MEMORY.md mid-session would stupidly invalidate your entire cache, forcing the AI to re-read everything. Now, it smartly processes only the modifications, preventing unnecessary cache invalidation.
  • Disk Space Defense: The background cleaner (Cron) that handles session debris now has a disk capacity limit (maxDiskBytes), preventing catastrophic server storage blowouts caused by endlessly accumulating logs.

3. 🤐 Assistant’s Lips Sealed: Telegram Stability & Reasoning Leak Prevention

Those subtle, annoying bugs when running the bot via messengers have been completely squashed.

  • Blocking Internal Monologue (<think>) Leaks: The recently popular Reasoning Models tend to grumble to themselves before answering. This update flawlessly blocks those raw reasoning processes from accidentally spilling into the chat window. Users only receive the clean, finalized response.
  • Improved Telegram Reactions: If the bot encounters an error while leaving an emoji reaction, the conversation no longer halts. Also, the dumb behavior of repeatedly reading messages when the bot restarts has been stopped.
  • Smart Error Recovery (Failover): When an API server goes down (HTTP 502/503, etc.), instead of stubbornly clinging to the dead server, the bot now smartly and quickly hops to an alternative fallback model to continue the conversation.

4. 🛡️ Fortress Mode: SSRF Default Changes and Security Hardening

Following version 2.22, the Security patches in this update are formidable.

  • Internal Network Access Blocked by Default (SSRF Defense): The most crucial change! The browser policy now blocks access to internal (Private) networks by default. The bot can no longer poke around your home/local network freely. (Of course, you can re-enable this in settings if needed.)
  • Command Obfuscation Blocking: When the bot executes system commands (Exec), it now possesses a sixth sense for hacker-style command obfuscation tricks. It strictly requires the owner’s explicit approval before running anything suspicious.
  • HTTP Security Headers: An option has been added to enforce strong security headers (HSTS) on the gateway Web UI.

Conclusion

While flashy new features are nice, for a tool like OpenClaw that touches the heart of my local system and personal data, “Security and Stability” must come first.

This 2.23 update is incredibly reassuring—cutting API costs with prompt caching, while ruthlessly blocking command obfuscation and unauthorized internal network access. I definitely need to set aside some time soon to test Kimi’s new video analysis feature!

Everyone, run the openclaw update command quickly to get on 2.23, and enjoy a safe and smooth AI assistant life!