OpenClaw v2026.4.20 Beta Brings Smarter Defaults and Major Plugin Upgrades

The OpenClaw team shipped a substantial pre-release yesterday with sweeping changes across models, plugins, and the onboarding experience.

Models & Defaults: Anthropic selections, Opus aliases, and bundled image understanding now default to Claude Opus 4.7. Moonshot/Kimi surfaces default to kimi-k2.6 with kimi-k2.5 preserved for compatibility. Moonshot thinking.keep = “all” is now supported on kimi-k2.6.

New Plugin Features:

  • Google Gemini TTS — text-to-speech support added to the bundled google plugin, including voice selection, WAV reply output, and PCM telephony output
  • Memory/LanceDB cloud storage — durable memory indexes can now run on remote object storage instead of local disk only
  • GitHub Copilot embeddings — a new embedding provider for memory search, with transport reuse for remote overrides, token refresh, and safer payload validation

Agent Improvements: The default system prompt and GPT-5 overlay received stronger completion bias, live-state checks, weak-result recovery, and verification-before-final guidance. A new experimental flag agents.defaults.experimental.localModelLean drops heavyweight default tools (browser, cron, message) for weaker local-model setups.

Setup Wizard Refresh: The security disclaimer got a yellow warning banner, section headings, bulleted checklists, and an un-dimmed note body. A loading spinner now prevents blank screens during the initial model catalog load.

Other highlights: Tiered model pricing support in cached catalogs, cron job state split into jobs-state.json for cleaner git tracking, BlueBubbles group system prompt forwarding with wildcard * fallback, and Mattermost now streams thinking and partial reply text into a single draft preview post.

Source: GitHub Release v2026.4.20-beta.1

← Back to News