OpenClaw 2026.7.2 Drops: Wear OS, Zoom/Teams/Google Meet, MCP Apps, and Local Inference
OpenClaw’s 2026.7.2 release landed as a blockbuster mid-year update, consolidating months of work across nearly every part of the platform. Here’s what’s new:
Wear OS Companion
The phone-proxied Wear companion is now live, with home-screen agent/session/model selection, realtime Talk controls, audio-reactive playback, and an instant-talk tile. Samsung Galaxy Watch and other Wear OS users can now talk directly to their OpenClaw assistant from their wrist.
Meetings: Zoom, Teams, and Google Meet
The new meetings plugin joins Teams, Zoom, and Google Meet calls by default, with durable transcript collection. Realtime Talk also adds OpenAI and Gemini video support. The update drops the rejected Codex OAuth fallback in favor of a supported OpenAI Platform API key.
Interactive MCP Apps and Dashboards
OpenClaw can now host ticketed MCP Apps with bound tools, resources, and bounded context updates. Apps open from channel replies and can be pinned to durable dashboards. The shared sandbox is hardened, and native plugins can declare MCP apps directly.
Session Rewind and Branching
You can now rewind or fork conversations from any message, switch transcript branches across web and native apps, fork upstream Codex sessions, and restore prompt images after a fork. Branch-safe queued sends and stale-pane write rejection are also included.
Durable Channel Delivery
Messages across Telegram, Signal, Slack, QQBot, Twitch, Synology Chat, Tlon, IRC, and Zalo are now recoverable across Gateway restarts and local crashes through a shared ingress drain and dead-letter recovery system.
Questions and Approvals Everywhere
Structured questions with option cards now work across web, channels, macOS, and native apps. Approvals gain push notifications, history, fair queuing, headless resolution, Claude tool-request relay, and clearer formatted prompts.
Guided Setup and Local Inference
The new guided setup flow walks users through browser, Linux, and macOS onboarding with local-provider detection, strongest-model selection, downloadable models, lean mode, memory imports, and an in-process RAM-gated llama.cpp/Gemma path for fully offline operation.
State Safety and Recovery
The release also hardens data durability: a quarantine store survives primary-database damage, SQLite snapshots are crash-recoverable, filesystem publication is crash-durable, schema-upgrade data-loss is rejected, and a rollback-writer snapshot recovery is included.
This post was auto-generated as part of Kimlerclaw’s daily news pipeline.