OpenClaw v2026.6.5: AI Agent Fixes for Thinking Leaks and MCP Handling

By FintechExtra
10 Jun 2026

OpenClaw just dropped version 2026.6.5, and it's a focused update that plugs two significant gaps in AI agent output handling. If you've been dealing with weird <thinking> tags appearing in your QQBot channels or hitting 400 errors when using MCP tools, this release is for you.

What Changed

The first fix targets QQBot. Previously, when the bot processed a model response that included reasoning scaffolding—those <thinking> tags—it would sometimes pass them straight into the channel. That's ugly. Worse, it leaks internal thinking. Thanks to @openperf's work, the bot now strips that content before delivery. The change is linked to issues #89913 and #90132.

The second change addresses MCP tool results. When a tool returns richer content like resource links, audio, or malformed images, the materialization boundary was failing to coerce these properly. That led to Anthropic-style 400 errors and poisoned session history. Now, those non-text blocks are handled gracefully. Credits go to @RanSHammer and @849261680 for the fix (issues #90710, #90728).

Why It Matters

These aren't flashy features. They're stability and reliability fixes that make OpenClaw a more robust foundation for AI agents. The thinking leak fix matters for trust—you don't want users seeing the raw thought process. It's a privacy and UX win. The MCP fix prevents cascading failures in complex tool chains. For enterprises deploying these workflows, that's critical. A single uncoerced block can break an entire session. Now it won't. This update quietly raises the bar for production-ready AI integrations.

Official Source: https://github.com/openclaw/openclaw/releases/tag/v2026.6.5