Voltar ao blog
Guia

OpenClaw 2026.2.23: Kilo Gateway, Vercel AI, and Smarter Session Management

24 de fevereiro de 2026 4 min de leitura
OpenClaw 2026.2.23

OpenClaw 2026.2.23 dropped on February 24th. It's a more focused release compared to the last two, but it brings some genuinely useful additions: new provider support, better session maintenance tooling, a Moonshot web search integration, and a round of fixes that make the agent more resilient across the board.

Kilo Gateway Support

The biggest new addition is first-class support for the Kilo Code gateway provider. If you're using Kilo Code, you can now connect OpenClaw directly with full auth, onboarding, model defaults, and prompt cache handling. The default model is kilocode/anthropic/claude-opus-4.6. It's a clean integration that should feel familiar to set up.

Vercel AI Gateway

OpenClaw now accepts Claude shorthand model references for the Vercel AI Gateway, like vercel-ai-gateway/claude-*. These get normalized to the correct Anthropic-routed model IDs automatically, so you don't have to fuss with exact model strings if you're routing through Vercel.

Moonshot Web Search and Video

The $web_search tool now supports Moonshot (Kimi) as a provider, with proper citation extraction from search results and a corrected two-step flow that echoes tool results before final synthesis. Moonshot also gets native video understanding support in this release, joining the existing audio and image handling. A nice expansion for anyone using Chinese AI providers.

Session Maintenance and Disk Budgets

OpenClaw now has a proper openclaw sessions cleanup command for managing session storage. You can set disk budget limits via session.maintenance.maxDiskBytes and highWaterBytes, and the cleanup handles transcript archives and run log retention safely. If you've been running OpenClaw for a while and noticed your session storage growing, this gives you real tools to manage it.

Prompt Cache Improvements

A couple of meaningful cache optimizations landed in this release. Bootstrap file snapshots (like AGENTS.md and MEMORY.md) are now cached per session key and cleared on reset, which reduces unnecessary prompt cache invalidations from in-session file writes. Per-agent params overrides also now properly merge on top of model defaults, so agents with different traffic patterns can tune their cache behavior independently.

Breaking Change: Browser SSRF Policy

One thing to be aware of before updating: the browser SSRF policy now defaults to trusted-network mode. The config key has also been renamed from browser.ssrfPolicy.allowPrivateNetwork to browser.ssrfPolicy.dangerouslyAllowPrivateNetwork. Running openclaw doctor --fix will migrate your config automatically, so this should be painless for most users.

Reasoning Behavior Fixes

A few fixes landed around reasoning output leaking into chat replies. When /reasoning off is active, raw reasoning text is now properly suppressed before delivery. When model-default thinking is enabled (like thinking=low), auto-reasoning stays disabled unless you explicitly turn it on. These were causing internal chain-of-thought text to occasionally show up in user-facing replies, which is now resolved.

Provider Reliability Improvements

Several provider-specific fixes made it into this release. HTTP 502, 503, and 504 errors are now treated as failover-eligible, so the agent can switch providers during upstream outages instead of retrying the same failing target. Groq throttling errors no longer incorrectly trigger overflow recovery. Bedrock prompt cache handling is now correctly scoped to Anthropic models only. And a fix for Anthropic OAuth tokens prevents 401 errors when params.context1m is enabled with subscription-based access.

Security Fixes

A couple of notable security improvements: OpenTelemetry diagnostics now redact API keys and tokens before export, so sensitive values don't end up in your observability pipeline. The openai-image-gen skill also had a stored XSS vulnerability in its HTML gallery generation that's now patched. The exec approval system also got better at detecting obfuscated commands before evaluating allowlist decisions.

How to Update

If you're running OpenClaw locally:

Code Bash
npm install -g openclaw@latest
openclaw doctor --fix
openclaw gateway restart

Note the --fix flag this release to handle the SSRF policy migration automatically.

If you're on ClawHosted, your instance will be updated automatically with no action needed on your end.

Links

  • Full Changelog on GitHub
  • OpenClaw Documentation
  • Community Discord
  • Running OpenClaw on ClawHosted? Updates like this one are applied automatically, so you're always on the latest version without touching the command line. Get started here.

    Por

    Renaud

    Clearly not an OpenClaw bot!

    Posts relacionados