OpenClaw 2026.4.22 adds local TUI, /models add, and Grok voice tools

• OpenClaw's release thread says v2026.4.22 adds a gateway-free local TUI, chat-time model registration, Tencent Hy3 support, and a broader xAI media surface.
• According to OpenClaw's /models add post, /models add <provider> <modelId> registers a model from chat with no Gateway restart, while OpenClaw's local TUI post says terminal chats can now run without a Gateway at all.
• OpenClaw's xAI tools post expands Grok support beyond text into image generation and editing, text-to-speech, batch speech-to-text, and Voice Call streaming transcription.
• For debugging, OpenClaw's diagnostics post describes a redacted bug-report export with logs and health snapshots but no prompts or secrets, while OpenClaw's trajectory bundles post adds a richer flight recorder for agent runs.
• The release also folds in token-saving plugin work: OpenClaw's tokenjuice note says openclaw plugins enable tokenjuice can reduce tool-call token use.

The official release notes are worth opening because the tweet thread is only the top layer. You can jump straight to the local TUI docs, the models-from-chat docs, the diagnostics page, and the trajectory bundle docs.

OpenClaw's most useful change is friction removal. The new local TUI runs terminal chats without a Gateway, but OpenClaw's local TUI post says plugin approval gates still stay in place.

Model registration also moves into the chat loop. In the models-from-chat docs, OpenClaw documents /models add <provider> <modelId> as an immediate registration path, matching OpenClaw's /models add post, which says the model becomes usable without restarting the Gateway.

That combination changes where setup happens:

• Local session first, via the local TUI docs
• Model catalog expansion from chat, per OpenClaw's /models add post
• Approval gates preserved even in local mode, per OpenClaw's local TUI post

The xAI integration is suddenly much wider than a plain text-model hookup. According to OpenClaw's xAI tools post and the matching xAI provider docs, OpenClaw now exposes:

• Grok image generation
• Grok image editing
• Text-to-speech
• Batch speech-to-text
• Voice Call streaming transcription

That puts xAI into the same conversation as the agent's other multimodal plumbing instead of leaving it as a single model endpoint.

OpenClaw split observability into two artifacts.

The first is diagnostics export. The diagnostics docs and OpenClaw's diagnostics post describe a redacted zip for bug reports that includes logs, health and status, config shape, and stability snapshots, but excludes prompts, chat text, tokens, and secrets.

The second is trajectory bundles. In the trajectory docs, OpenClaw describes a flight recorder for agent runs that packages redacted transcripts, runtime events, prompts, metadata, and artifacts for reproducible debugging and dataset export, which matches OpenClaw's trajectory bundles post.

The interesting split is that diagnostics export is for host health, while trajectory bundles are for replaying what the agent actually did.

Tencent Cloud arrives through TokenHub rather than as a bare provider stub. OpenClaw's Tencent Hy3 post says Hy3 ships as a preview entry in the model catalog with onboarding and pricing metadata, and the Tencent provider docs show the same packaging.

The other quiet addition is tokenjuice. OpenClaw's tokenjuice note says openclaw plugins enable tokenjuice can cut tool-call token usage, and vincent_koc's tokenjuice release post adds that tokenjuice 0.6.1 is bundled in the next OpenClaw release with more compaction rules, Windows improvements, and better doctor and stats output.

For a low-key release, that is a lot of surface area: local execution, live model registration, multimodal xAI tools, two debugging artifacts, a new Tencent path, and an efficiency plugin all landed in one cut.