OpenClaw 2026.3.28: Approval Gates for Tool Calls and Built-In xAI Search

OpenClaw 2026.3.28 introduced a set of changes that matter for people running self-hosted/personal agents in production.

What changed (and why it matters)

1) Approval gates can now happen before a tool call runs

The release adds async approval hooks for tool execution. In practice, this gives you a stronger safety checkpoint for high-risk actions (shell, network writes, external sends) before they execute.

Why this is practical:

• You can keep agents autonomous for low-risk tools
• Require human approval for sensitive operations
• Apply one policy pattern across Telegram/Discord/UI approval surfaces

2) xAI integration moved to Responses API + first-class x_search

OpenClaw’s bundled xAI provider now uses the Responses API path and includes native x_search support with onboarding/config wiring.

Why this is practical:

• Lower setup friction for teams already using Grok/xAI
• Cleaner web-search enablement in mixed-provider stacks
• Fewer manual plugin toggles during initial setup

3) Config migration behavior tightened

Very old legacy config keys are no longer silently rewritten forever. If they are too old, they now fail validation and require explicit cleanup.

Why this is practical:

• Fewer “mystery” config mutations
• Clearer upgrade path for long-lived self-hosted installs
• Better drift control in infra-as-code environments

Suggested operator checklist

If you run OpenClaw for real users, do this after upgrading:

1. Define approval policy by tool risk (auto-allow, require approval, deny).
2. Test approval UX in your primary channel (Telegram/Discord/web UI) with one harmless command and one sensitive command.
3. Verify xAI search behavior if you use Grok/xAI in production flows.
4. Run config validation/doctor and clean up any legacy keys that now fail fast.
5. Document rollback steps before broad rollout across multiple instances.

Source cross-check

• Official release stream: GitHub Releases for openclaw/openclaw
• Independent mirrors/trackers: ReleaseBot and PatchBot summaries for 2026.3.28

If you’re running personal agents with mixed autonomy and human-in-the-loop controls, this release is worth prioritizing because it improves both safety boundaries and day-2 operability.