OpenClaw v2026.4.22 dropped overnight and it's one of the more feature-dense releases in recent memory. The headline is full xAI provider support — Grok now handles image generation, text-to-speech, and speech-to-text alongside its existing chat capabilities. But there are several other changes worth knowing about, especially if you run voice pipelines or a gateway-less setup.
Here's everything that changed and what it means for your setup.
1. xAI / Grok Gets Full Multimodal Support
Previously, xAI in OpenClaw meant Grok for chat. v2026.4.22 turns it into a full multimodal provider:
- Image generation:
grok-imagine-imageandgrok-imagine-image-pro— with reference-image edit support (send an existing image, get a modified version back) - Text-to-speech: Six live xAI voices, MP3/WAV/PCM/G.711 output formats
- Speech-to-text:
grok-sttfor batch audio transcription; xAI realtime transcription for Voice Call streaming
This makes xAI a serious competitor to OpenAI's full provider stack within OpenClaw. If you have an xAI API key and want to stay on one provider for text + voice + images, you now can.
How to Configure xAI Image Gen
# In openclaw.json → providers → xai
{
"providers": {
"xai": {
"apiKey": "your-xai-key",
"imageModel": "grok-imagine-image",
"ttsVoice": "aurora"
}
}
}
Voices available: aurora, echo, flash, nimbus, orion, zephyr (exact names may vary — check /models after update).
2. Streaming STT for Deepgram, ElevenLabs, and Mistral
Real-time Voice Call transcription now extends beyond OpenAI and xAI. Three more providers gain streaming STT in this release:
- Deepgram — Voice Call streaming transcription
- ElevenLabs — Voice Call streaming + Scribe v2 batch audio transcription for inbound media
- Mistral — Voice Call streaming transcription
If you're running OpenClaw for voice-first workflows (call summaries, live transcription, voice-driven task entry), this significantly expands your provider options beyond OpenAI's realtime API.
ElevenLabs users: Scribe v2 batch transcription is new — it processes inbound audio media (voice messages on WhatsApp, Telegram, etc.) rather than just live streams. This makes ElevenLabs a solid all-in-one voice stack alongside its TTS capabilities.
3. Local TUI Mode — Gateway No Longer Required for Terminal Chat
This is a quiet but meaningful change. Previously, OpenClaw's terminal UI (TUI) required a running gateway daemon. v2026.4.22 adds a local embedded mode that lets you run terminal chats without a gateway — while still enforcing plugin approval gates.
What this means:
- You can now use OpenClaw from the terminal on a machine that doesn't have the gateway running
- Useful for portable setups, air-gapped machines, or quick one-off chats without spinning up the full daemon
- Plugin approval gates still apply — you're not bypassing security, just skipping the daemon dependency
openclaw chat --local # runs embedded, no gateway needed
Note: Local mode won't have access to channel-specific context (your Telegram/WhatsApp conversations, etc.) since those route through the gateway. It's best for quick tasks, testing, and offline use.
4. /models add — Register Models From Chat
Previously, adding a custom model required editing openclaw.json and restarting the gateway. v2026.4.22 adds a /models add command that registers a model live from chat:
/models add openai gpt-5.5
/models add xai grok-imagine-image-pro
/models add ollama llama3.3:70b
No restart needed. The model is available immediately after the command. The existing /models command (provider browser) is unchanged — /models add is an extension, not a replacement.
5. WhatsApp Native Reply Quoting
OpenClaw can now quote messages in WhatsApp replies natively, using configurable replyToMode. This means when your agent responds to a specific WhatsApp message, it shows as a threaded reply rather than a standalone message — which is how humans actually communicate in WhatsApp.
Also new for WhatsApp: per-group and per-direct system prompt overrides. You can now configure different agent personalities for different WhatsApp groups or DM conversations:
# openclaw.json example
"channels": {
"whatsapp": {
"groups": {
"Family Chat": { "systemPrompt": "You are a friendly family assistant..." },
"Work Team": { "systemPrompt": "You are a professional work assistant..." },
"*": { "systemPrompt": "Default prompt for all other groups" }
}
}
}
6. Onboarding Auto-Install
A minor but welcome fix: missing provider and channel plugins now auto-install during first-run setup. Previously, if you skipped a plugin step during onboarding, you'd hit cryptic errors later and need to manually recover. Now setup just handles it.
7. Bug Fix: Plugin Discovery Hang (lossless-claw compat)
The strict-match contract added in v2026.4.14 broke plugins whose internal engine id doesn't match their registered slot id — most notably lossless-claw and similar third-party plugins. This produced repeated info.id must match registered id errors on every turn.
v2026.4.22 fixes this. If you've been seeing that error since mid-April, update and it should clear.
Full Change Summary
- NewxAI: image generation (grok-imagine-image / grok-imagine-image-pro), 6 TTS voices, grok-stt transcription, realtime STT streaming
- NewSTT: Deepgram, ElevenLabs (+ Scribe v2 batch), Mistral all gain Voice Call streaming transcription
- NewTUI: local embedded mode — run terminal chat without a gateway (plugin gates still enforced)
- New/models add <provider> <modelId> — register models live from chat without restart
- NewWhatsApp: configurable native reply quoting (replyToMode); per-group and per-direct system prompt overrides with wildcard fallback
- NewOnboarding: auto-install missing plugins during first-run setup
- NewOpenAI/Responses: native web_search auto-used for direct Responses models when search enabled
- NewSessions: mailbox-style sessions_list filters (label, agent, search)
- FixPlugin discovery hang — lossless-claw and plugins where engine id ≠ slot id now work correctly
- FixAgents/compaction: Pi compaction events renamed to compaction_start/compaction_end to align with pi-coding-agent 0.66.1
- FixModels/auth: provider-owned default-model additions now merge instead of replace on re-auth
How to Update
Update command
npm update -g openclaw && openclaw gateway restart
Check your version after: openclaw --version should show 2026.4.22 or later.
If you were affected by the lossless-claw plugin hang: After updating, do a full gateway restart (openclaw gateway stop && openclaw gateway start) rather than just a reload. The plugin registry needs to be rebuilt cleanly.
What This Release Means for Your Setup
If you're running a voice-first setup, this is a significant upgrade — Deepgram, ElevenLabs, and Mistral users can now use streaming STT for Voice Call without depending on OpenAI's realtime API. If you're a Grok user, you now have a complete multimodal stack in one provider.
The local TUI mode and /models add command are quality-of-life wins that reduce friction in daily use — especially for operators who frequently test new models or use OpenClaw from multiple machines.
Running an Older OpenClaw Version?
Staying current matters — especially with active security CVEs in the ecosystem right now. ClawReady's $49 audit covers your version, config, and exposure — and flags anything that needs attention before it becomes a problem.
Get a Security Audit →