OpenClaw, the open-source AI assistant project, just dropped version v2026.5.22 with some serious performance tuning and new plugin support. The headline feat? By pre-warming service provider authentication states at gateway startup, they've crushed the model list fetch latency from a sluggish 20 seconds down to just 5 milliseconds — that's roughly a 4100x improvement. This pre-warming trick lets hot-path model list requests skip the usual messy plugin and CLI discovery flow entirely.
On the plugin front, the new release introduces a standalone meeting-notes external plugin (separate from the core npm package) with Discord voice channels as its first real-time input source. Users can either access recordings through the read-only CLI tool openclaw meeting-notes or manually import them as text. To keep multi-agent deployments from ballooning context windows and leaking sensitive info, the team also reworked the sub-agent spawning context. Now, by default, only AGENTS.md and TOOLS.md are passed down to delegated sub-agents — personal settings, user memory, and identity key files are locked away.
Under the hood, the project bumped the protobufjs dependency to 8.4.0 to patch a security hole. For xAI integration, the new version can reuse Grok's OAuth authorization file for web searches. And OpenAI image generation has been rerouted to the native image API, finally squashing the bug where legitimate API keys would get 401 errors from the Codex streaming backend.