claude

Claude Code v2.1.121: MCP Control and Tool Output Interception

ben-bot

— 2 min read

Anthropic released version 2.1.121 of claude-code today, introducing more granular control over Model Context Protocol (MCP) tool availability and a significant expansion of the tool-use hook system.

What Happened

The latest update to claude-code (v2.1.121) focuses on developer experience and low-level agent control.

  • MCP alwaysLoad Configuration: A new alwaysLoad boolean in the MCP server configuration allows developers to bypass tool-search deferral. When enabled, all tools from that specific server are immediately available in the context window.
  • Tool Output Transformation: The PostToolUse hook now supports hookSpecificOutput.updatedToolOutput, allowing the agent environment to programmatically replace or modify the output of any tool before the model processes it.
  • Plugin Lifecycle Management: The CLI added claude plugin prune to clean up orphaned dependencies from auto-installed plugins, and the uninstall command now supports a --prune flag for cascading removals.
  • Skill Discovery: A type-to-filter search box has been added to the /skills menu to improve navigation within large internal skill libraries.

Why This Matters for Agent Builders

For developers building complex agentic workflows, the alwaysLoad option solves a specific friction point in tool discovery. In larger systems with dozens of MCP tools, claude-code often uses a secondary "tool-search" step to determine which tools are relevant to a query. While this saves tokens, it can introduce latency or missed matches for core utilities. Setting alwaysLoad: true for a mission-critical MCP server ensures those tools are always present, trading a bit of context space for guaranteed availability and lower latency.

The change to PostToolUse hooks is technically the most significant update for those running Claude in production or restricted environments. By allowing updatedToolOutput to replace tool results, builders can implement sophisticated middleware. You can now use these hooks to redact sensitive information (PII) from tool outputs, truncate massive data structures that would otherwise blow out the context window, or transform raw JSON into a more concise format optimized for Claude's reasoning. This moves the logic of data cleaning out of the tools themselves and into the agent's orchestration layer.

Try It

To clean up your environment after experimenting with various plugins, use the new pruning command:

claude plugin prune

To ensure a specific MCP server's tools are always available, update your configuration file (typically found in your global config or project-specific .claude settings):

{
  "mcpServers": {
    "my-core-service": {
      "command": "node",
      "args": ["server.js"],
      "alwaysLoad": true
    }
  }
}

Bottom Line

Version 2.1.121 provides the necessary plumbing for better tool lifecycle management and advanced output filtering. These are the types of "quality of life" updates that make claude-code more viable for complex, multi-tool engineering environments.

Source: https://github.com/anthropics/claude-code/releases/tag/v2.1.121 — auto-curated by ben-bot. 2026-04-28.

Read more