OpenClaw in March: Tool Gates, Grok Search, Image Gen, and Why It Matters

March was a big month for OpenClaw. Three releases shipped (3.23, 3.24, 3.28), each adding real capabilities — not cosmetic changes. Here's what matters for Agentbot users and why your agent just got significantly more capable.

1. Your Agent Now Asks Before It Acts

The biggest feature in 3.28: Tool Approval Gates. Before this, your agent would execute tools — shell commands, API calls, file writes — without asking. Now it can pause and ask for your permission first.

When a tool call hits an approval gate, you'll see a prompt in your channel (Telegram, Discord, webchat) asking you to approve or deny. One tap to allow, one tap to block. The agent waits for your answer before proceeding.

This changes the trust model entirely. Your agent has the capability to do dangerous things — delete files, send messages, move money. Now you decide when it crosses that line, not a hardcoded safety rule.

2. Grok Gets Native Web Search

If your agent uses xAI's Grok models, it now has first-class web search viax_search. No manual configuration — the xAI plugin auto-enables from your existing web search and tool config.

Grok's search is different from Tavily or Perplexity. It's built into the model's reasoning pipeline, so search results integrate more naturally with the agent's thought process. Better answers, fewer hallucinations on current events.

3. MiniMax Image Generation

image-01is now available as a MiniMax image generation provider. Text-to-image and image-to-image editing with aspect ratio control. Your agent can now generate images on demand — no external API keys needed if you're on a supported plan.

4. ACP Channel Binds

Want to run a Codex or Claude session directly in your current chat? Now you can. /acp spawn codex --bind hereturns your active conversation into a coding workspace — no child threads, no context loss. Works on Discord, iMessage, and BlueBubbles.

This is a game changer for developers. Instead of switching between chat and a separate coding tool, your agent IS the coding tool. Right there in your conversation.

5. Smarter Rate Limiting

Rate limit cooldowns are now scoped per model. Before, one 429 error on any model would block every model on the same auth profile. Now, if GPT-4o rate limits, your agent can still use Claude or Gemini without waiting. The cooldown ladder is stepped: 30s → 1m → 5m instead of exponential escalation.

What This Means for Agentbot Users

Every OpenClaw improvement ships automatically to all Agentbot agents. You don't need to update, configure, or do anything. Your agent gets smarter the moment we deploy.

The theme across these releases: your agent is becoming more capable and more controllable at the same time. More tools, more models, more integrations — but also more ways for you to set boundaries and approve actions. That's the right balance.

We're not building an agent that does everything automatically. We're building an agent that CAN do everything, and asks you first when it matters.

The Numbers

• 80+ bug fixes across platforms (Telegram, Discord, WhatsApp, iMessage, Matrix, Slack)
• 5 new tool providers (Grok search, MiniMax image, ACP binds, CLI backends)
• 3 releases in March (3.23, 3.24, 3.28)
• 12+ channel integrations now stable
• 34 models available via smart routing

OpenClaw v2026.3.28 · Running on all Agentbot agents · No action needed