Claude Code's latest update packs several features that deepen the tool's agentic capabilities and give developers finer control over how the model thinks.
The /effort command is deceptively simple but immediately useful. It lets you set the model's thinking depth without restarting your session — dial it down for quick tasks, crank it up when you need deep architectural reasoning. Combined with the 'ultrathink' keyword for maximum effort, developers now have granular control over the trade-off between speed and depth on a per-prompt basis.
MCP Channels (research preview) are the more significant architectural addition. With the --channels flag, MCP servers can push messages directly into your Claude Code session. This enables real-time monitoring integrations — imagine your CI server pushing build results, your error tracker sending alerts, or your deployment pipeline streaming status updates directly into your coding session without you needing to switch context.
Starting from v2.1.76, MCP servers can also request structured input during task execution. Instead of interrupting the workflow with a text prompt, the server displays an interactive form or opens a URL to collect the necessary data. This is a meaningful step towards smoother human-in-the-loop agentic workflows where the agent can gather exactly the information it needs without breaking flow.
Other improvements include voice mode now supporting 10 additional languages (Russian, Polish, Turkish, Dutch, Ukrainian, Greek, Czech, Danish, Swedish, Norwegian) and a fix for compound bash command permissions that was causing repeated approval prompts.
