OpenClaw 4.1 Beta vs 3.31: MCP HTTP Transport, Cross-Agent Memory, and What the Channel Sprawl Means

Tools

OpenClaw 4.1 Beta Adds Remote MCP HTTP Transport and Per-Server Connection Timeouts

MCP servers no longer have to run as local stdio processes. You can now point mcp.servers at a remote HTTP or SSE URL with auth headers, which is a real unlock for hosted or multi-machine setups. Per-server connection timeouts also mean flaky remote MCP endpoints stop hanging your entire session.

4.1 Beta Fixes Gateway Task Registry Stall That Was Hanging Upgraded Gateways at Startup

A critical bug where the task registry maintenance sweep was blocking the gateway event loop under SQLite pressure caused gateways to hang about a minute after startup. If you upgraded to 3.31 and noticed your gateway freezing post-launch, this is the fix.

Cross-Agent Memory Search Lands in 3.31 with Per-Agent extraCollections Config

Agents can now opt into searching other agents' session transcripts via memorySearch.qmd.extraCollections without collapsing everything into one shared namespace. For multi-agent setups running a researcher plus a coder plus an orchestrator, this is the first real building block for agents that actually share context.

Bundled MCP Tools Now Use serverName__toolName Namespacing to Kill Collision Bugs

Tool name collisions between MCP servers that both expose a generic name like search or fetch are now resolved by materializing names as serverName__toolName. Sounds minor until you've spent an afternoon debugging the wrong tool getting called silently.

4.2 before_agent_reply Hook Lets Plugins Short-Circuit the LLM Entirely

Plugins can now intercept incoming messages and return a synthetic reply before the model is ever called. For high-traffic channels where a large chunk of messages are routine commands or FAQ hits, this hook alone could cut LLM costs significantly by handling them at the plugin layer.

Analysis

OpenClaw 3.31 Operator Breakdown: What the Fail-Closed Install Scanner Means for Community Skills

Plugin and skill installs now fail by default when dangerous-code findings are detected, requiring an explicit --dangerously-force-unsafe-install override. Community skills on ClawHub that previously installed fine may start failing until authors clean up flagged patterns, so audit your installed skills before upgrading.

Obfuscation Detection in 3.31 Overrides ask:off and Blocks Legitimate Agent Commands

The new detectCommandObfuscation function in 3.31 flags heredocs, base64 payloads, eval, curl-pipe-shell, and commands over 10k characters as suspicious, then forces approval regardless of your ask setting. AI agents routinely generate commands that trip all of these, so if your agent is getting blocked on normal operations after upgrading, this hardcoded override is why.