OpenClaw 4.1: Task Brain, SearXNG, and Bedrock Guardrails | OpenClaw DC

OpenClaw v2026.4.1 is here, released today. The headline feature is Task Brain, a chat-native background task board that acts as an OS-level control panel for your AI agent. Task Brain lets the agent schedule its own work, track progress across conversations, and refuse commands that violate trust boundaries you define. Beyond Task Brain, this release bundles a SearXNG provider plugin for the web_search tool, adds Amazon Bedrock Guardrails support for content filtering, introduces macOS Voice Wake for Talk Mode, and delivers reliability fixes across webchat, Telegram, Discord, and session handling. A total of 104 contributors rewrote underlying code to make this release possible.

How to Upgrade

One command:

npm install -g openclaw@latest

Verify: openclaw --version should show 2026.4.1.

If you hit problems, check the troubleshooting guide. Coming from 3.24 or earlier? Read the 3.24 release notes first so you do not miss changes to sub-agent configuration and the skills UI.

Task Brain: Chat-Native Background Task Board

Task Brain is the biggest feature in 4.1 and the result of a ground-up rewrite of how OpenClaw handles persistent work. Previous versions treated every conversation as a standalone interaction. Task Brain changes that by giving the agent a dedicated task board it can read, write, and schedule against.

Access it from any chat interface with the /tasks command. You see a board of active, queued, and completed tasks. The agent can create tasks on its own when a request requires multiple steps, or you can add them manually. Each task carries a status, a priority level, a deadline, and a log of every action the agent took to complete it.

The real power is in the trust boundary enforcement. You define rules in the config that tell the agent what it is allowed to do autonomously and what requires human approval:

openclaw config set tasks.boundaries.autoApprove '["file_read","web_search","summarize"]'
openclaw config set tasks.boundaries.requireApproval '["file_write","api_call","purchase"]'
openclaw config set tasks.boundaries.deny '["delete_database","sudo"]'

When the agent encounters a task step that falls into the requireApproval list, it pauses and notifies you through your configured channel. If a step falls into the deny list, the agent refuses outright and logs the attempt. This gives you a structured way to let the agent operate independently on low-risk work while keeping humans in the loop for anything consequential.

Task Brain also supports scheduling. The agent can queue a task to run at a specific time or on a recurring interval. A nightly report generation task, a weekly data backup check, or a daily inbox triage all run from the same board. Scheduled tasks respect the same trust boundaries as interactive ones.

Bundled SearXNG Provider Plugin

OpenClaw 4.1 ships with a SearXNG provider plugin that maps directly to the web_search tool. SearXNG is a free, self-hosted metasearch engine that aggregates results from multiple sources without requiring API keys from Google, Bing, or any other provider.

To enable it, point the plugin at your SearXNG instance:

openclaw config set providers.searxng.url http://localhost:8888
openclaw config set providers.searxng.engines '["google","duckduckgo","brave"]'

Once configured, any skill or agent that calls web_search routes through your SearXNG instance automatically. Results come back as structured JSON with titles, URLs, and snippets. The agent uses them like any other tool output.

This is part of a broader push toward bundled provider and plugin support. Instead of requiring users to find, install, and configure third-party plugins for basic functionality, OpenClaw now ships common integrations out of the box. SearXNG is the first bundled provider plugin, and more are planned for upcoming releases.

Amazon Bedrock Guardrails

Enterprise users running OpenClaw with Amazon Bedrock models can now attach Bedrock Guardrails directly to the agent pipeline. Guardrails filter both input and output for sensitive content, PII, toxic language, and custom policy violations.

Configuration ties into your existing Bedrock setup:

openclaw config set bedrock.guardrailId YOUR_GUARDRAIL_ID
openclaw config set bedrock.guardrailVersion DRAFT

When a guardrail triggers, the agent receives a filtered response and logs the intervention. You can review guardrail events in the dashboard under the Bedrock activity tab. This pairs well with Task Brain’s trust boundaries, giving you two layers of control: one at the task level and one at the model output level.

macOS Voice Wake for Talk Mode

Talk Mode, the voice-driven interface for OpenClaw, now supports a wake word on macOS. Say “Hey Claw” and the agent starts listening without requiring a keyboard shortcut or mouse click.

Enable it with:

openclaw config set voice.wake.enabled true
openclaw config set voice.wake.phrase "Hey Claw"

The wake word detection runs locally using Apple’s Speech framework. No audio leaves your machine until the wake word is detected and you begin speaking your command. You can customize the wake phrase to any short keyword or name that works for your environment.

Voice Wake requires macOS 14 or later and a microphone with reasonable ambient noise levels. It works alongside existing Talk Mode hotkeys, so you can use either trigger method depending on the situation.

Reliability Fixes

Version 4.1 includes targeted fixes across several platform integrations:

• Webchat: Resolved a race condition that caused duplicate messages when users sent rapid follow-ups during streaming responses.
• Telegram: Fixed session persistence so conversations resume correctly after the bot restarts. Previously, a restart could orphan active sessions.
• Discord: Corrected a threading bug where replies landed in the wrong channel when multiple threads were active simultaneously.
• Sessions: Improved memory cleanup for long-running sessions that exceeded the context window. The agent now summarizes and archives context instead of silently dropping older messages.
• Approvals and failover: Strengthened the approval flow so that pending approvals survive a server restart. Failover between provider models now retries with exponential backoff instead of failing immediately.

These fixes address issues reported by the community over the past two release cycles. If you experienced any of these problems on 3.23 or 3.24, upgrading resolves them.

Should You Upgrade?

If you want autonomous background tasks: Yes. Task Brain is a foundational feature that changes how you interact with OpenClaw. Scheduled tasks, trust boundaries, and persistent progress tracking open up workflows that were not possible before.

If you self-host SearXNG: Yes. The bundled provider plugin eliminates the need for a separate web search API subscription.

If you run Bedrock models in production: Yes. Native Guardrails support adds enterprise content filtering without third-party middleware.

If you use Talk Mode on macOS: Yes. Voice Wake removes the friction of reaching for a hotkey.

If you rely on webchat, Telegram, or Discord: Yes. The reliability fixes resolve real bugs that affected daily usage.

If you are stable on 3.24 and none of these features apply: Task Brain alone is worth evaluating. It is the kind of feature that changes how you think about what an agent can do.

Questions or Need Help Setting Up?

Book a Call. We walk through upgrades, Task Brain configuration, SearXNG setup, and Bedrock Guardrails regularly with OpenClaw users.