ClawHosters
by Daniel Samer
3 min read
22 pull requests in four days. The OpenClaw v2026.2.21 release shipped on February 21, and the maintainers described it as a "make everything better" release. That's vague. So here is what actually changed.
Discord Voice Channel Bot: Talk to Your AI
This is the big one. The /vc join command turns your Discord AI chatbot into a voice participant. It joins a voice channel, listens through Voice Activity Detection, transcribes with OpenAI Whisper or Deepgram (streaming mode cuts roughly a second of latency), and responds via TTS through OpenAI, ElevenLabs, or Kokoro for offline setups.
Barge-in works. You can interrupt the bot mid-sentence and it stops talking. Auto-reconnect is built in. And if your server goes offline for maintenance, the bot picks the channel back up when it comes back.
For anyone running a community Discord server with an AI Discord bot, this changes how people interact with it. Voice is more natural than typing, especially for quick questions.
Gemini 3.1 Pro Preview
Google released Gemini 3.1 Pro Preview on February 19. OpenClaw merged support within two days. The benchmarks are hard to ignore: 80.6% on SWE-Bench, 94.3% on GPQA Diamond, and a score of 57 on the Artificial Analysis Intelligence Index (the average sits around 26). Context window is 1M tokens. Pricing runs $2/$12 per million tokens.
If you have been running Opus for everything and the bills make you wince, this is probably worth testing.
Volcano Engine and BytePlus (Doubao)
ByteDance launched Doubao-Seed 2.0 on February 14, and OpenClaw now supports it through both Volcano Engine (China) and BytePlus (international). There is also Doubao-Seed-Code, which targets agentic coding workflows. Both are Anthropic API compatible, so your existing LLM config works with minimal changes.
The pricing is roughly one-tenth of Western frontier models. Quality varies by task, but for high-volume use cases where you are watching your token costs, it is another option in the mix.
Per-Channel Model Overrides
channels.modelByChannel lets you map specific channel IDs to specific models. Run Opus on your Telegram AI bot for deep conversations. Run Sonnet on Discord for faster group chat responses. Run a free model on your public channel to keep costs zero. The config is straightforward and does not require separate gateway instances.
OpenClaw v2026.2.21 Browser Sandbox Security
The browser automation sandbox got a security pass. Dedicated Docker networking, CDP ingress CIDR restrictions, mandatory VNC passwords, and SHA-256 replacing SHA-1 for hash verification. The Pi runner retry loop is now capped at 32 to 160 attempts, which prevents runaway loops on flaky connections. Details in the security overview.
On ClawHosters, your instance is already running v2026.2.21. If you have not tried the free trial yet, this is a good release to start with.
