The new version of OpenClaw reduces the model list loading time by 4,000x and adds support for recording Discord voice calls.

According to Perceive Beating monitoring, the open-source AI assistant project OpenClaw has released version v2026.5.22, bringing significant updates to gateway performance and external plugin support. In this new version, by preheating service provider certification status at gateway startup, the response latency of the model list retrieval interface has been reduced from about 20 seconds to 5 milliseconds, achieving a nearly 4100x performance leap. The preheating state mechanism allows model listing requests in the hot path to bypass the cumbersome plugin and external command-line discovery process directly.

On the plugin ecosystem front, the new release introduces for the first time an external plugin for meeting notes separate from the core npm package, with support for Discord voice channels as the first real-time input source. Users can access recordings through a read-only command-line tool `openclaw meeting-notes` or manually import recordings as text. To reduce context overhead in multi-agent deployments and mitigate sensitive information leaks, the new version has also refactored the startup context mechanism for sub-agents, now by default transmitting only `AGENTS.md` and `TOOLS.md` to delegated sub-agents, completely isolating personal settings, user memories, and identity key files.

In terms of underlying dependencies and ecosystem integration, the project has upgraded the protobufjs dependency to version 8.4.0 to address security vulnerabilities. For xAI integration, the new version supports reusing Grok's OAuth authorization file for web searches. Additionally, the new version reroutes OpenAI image generation to the native image interface, completely resolving the historical flaw of legitimate API keys causing a 401 error due to authentication failure in the Codex streaming backend.