Claude Code 2.1: How MCP Tool Search Changed Everything

TL;DR: Claude Code 2.1 introduces MCP Tool Search, which automatically defers tool loading when descriptions exceed 10% of context. Result: 85% token reduction, accuracy jumping from 79.5% to 88.1% on Opus 4.5. Combined with the new hooks system and skills hot reload, this is the biggest productivity upgrade since launch.

What Problem Does MCP Tool Search Actually Solve?

If you've connected more than a few MCP servers to Claude Code, you've hit the wall. I had 7 servers running—GitHub, Slack, memory, sequential thinking, a couple custom ones—and my context was nearly exhausted before I typed a single prompt.

The numbers are brutal:

• 5 typical MCP servers (GitHub, Slack, Sentry, Grafana, Splunk) = ~55K tokens consumed
• One user reported 144,802 tokens from MCP tools alone
• MCP_DOCKER alone consumed 125,964 tokens across 135 tools

Thariq Shihipar, who announced the feature, put it plainly: "Users were documenting setups with 7+ servers consuming 67k+ tokens."

That's context pollution at scale. And Claude Code 2.1 fixes it.

How MCP Tool Search Works

Instead of loading every tool definition upfront, Claude Code now implements lazy loading:

1. Detection: Claude Code checks if MCP tool descriptions would use more than 10% of context
2. Deferral: When triggered, tools are loaded via search instead of preloaded
3. Discovery: Claude searches for and loads only the tools it needs, when it needs them

The feature supports two search modes:

The Performance Gains

Anthropic's benchmarks tell the story:

This isn't incremental. It's a step change.

What's New in Claude Code 2.1

Version 2.1.0 shipped January 7, 2026. Version 2.1.9 followed with 109 CLI refinements. Across the 2.1.x series, the team shipped 1,096 commits. Here's what matters:

MCP Tool Search (Now Default)

You don't need to opt in. It's enabled automatically for all users. When your MCP tool descriptions exceed the 10% threshold, Claude Code switches to search-based discovery.

If you want to force it on earlier or adjust behavior:

ENABLE_TOOL_SEARCH=true

Hooks System: PreToolUse and PostToolUse

Hooks let you inject custom logic at key points in Claude's workflow:

The power move: PreToolUse hooks can modify tool inputs before execution (starting in v2.0.10). Instead of blocking Claude and forcing retries, you intercept and correct.

Example use cases:

• Auto-add --dry-run flags to dangerous commands
• Redact secrets before they hit the terminal
• Enforce commit message formatting
• Auto-install dependencies before builds

Skills Hot Reload

Previously, creating or updating a skill required restarting your session. Now:

• Skills in ~/.claude/skills or .claude/skills are immediately available without restart
• Nested directories are auto-discovered—if you're editing packages/frontend/app.tsx, Claude finds skills in packages/frontend/.claude/skills/
• Hooks in skill frontmatter: Define PreToolUse, PostToolUse, and Stop hooks directly in your skill files

This enables real monorepo workflows where each package can have its own specialized skills.

Additional Features Worth Knowing

• Shift+Enter for newlines: No more fighting with the terminal
• Wildcard tool permissions: Bash(*-h*) grants help flag access across all commands
• /teleport to Claude.ai: Move your session to the web interface mid-conversation
• Multilingual output: Configure Claude to respond in Japanese, Spanish, etc.
• Background tasks: Ctrl+B to background long-running operations

How Should AI Architects Think About This?

MCP Tool Search isn't just a performance optimization. It changes how you can architect AI systems.

Before: Careful Tool Curation

You had to choose which MCP servers to connect, knowing each one consumed context. Complex setups meant tradeoffs.

After: Connect Everything

With lazy loading, you can connect every server you might need. Claude discovers what's relevant. The constraint shifts from "what can I afford to load" to "what might be useful."

For enterprise deployments, this means:

• Richer tool libraries without context penalties
• Domain-specific MCP servers can coexist without conflict
• Agentic workflows with dozens of tools become practical

The Hooks System Changes Security Posture

PreToolUse hooks with input modification enable:

• Transparent sandboxing: Dangerous commands get dry-run flags automatically
• Secret protection: Environment variables redacted before execution
• Compliance enforcement: All file writes go through review

This is how you build guardrails for AI coding assistants without breaking the developer experience.

What About Accuracy Concerns?

Fair question. Third-party testing shows mixed results at extreme scale:

The takeaway: MCP Tool Search works excellently for typical setups (5-50 tools). At thousands of tools, you'll want dedicated optimization like Stacklok's MCP Optimizer.

For most developers and architects, the default behavior is a massive win.

Getting Started

Verify Your Version

claude --version
# Should be 2.1.x or higher

Upgrade If Needed

npm install -g @anthropic-ai/claude-code@latest
# or
claude update

Check MCP Status

claude mcp list

Configure Hooks (Optional)

Create ~/.claude/settings.json:

{
  "hooks": {
    "PreToolUse": {
      "Bash": {
        "command": "your-validation-script.sh"
      }
    },
    "PostToolUse": {
      "Write": {
        "command": "prettier --write $FILE_PATH"
      }
    }
  }
}

Frequently Asked Questions

Is MCP Tool Search enabled by default?

Yes. As of January 2026, it's on for all users automatically. It triggers when tool descriptions exceed 10% of context.

Does Tool Search work with all MCP servers?

Yes. It works with any MCP server. The search happens client-side in Claude Code.

Can I disable Tool Search?

You can, but why would you? If you have specific needs, check Claude Code settings.

What's the difference between Regex and BM25 search modes?

Regex is precise pattern matching (get_.*_data). BM25 is semantic search with natural language queries. Claude picks the appropriate mode automatically.

Do hooks slow down Claude Code?

The default hook timeout changed from 60 seconds to 10 minutes in 2.1.x, but well-written hooks add negligible overhead. Keep them fast.

Does skills hot reload work in all environments?

Yes. Works on macOS, Linux, and Windows. Skills are discovered from both global (~/.claude/skills) and project-local (.claude/skills) directories.

Key Takeaways

1. MCP Tool Search cuts token usage by 85% while improving accuracy
2. Hooks enable sophisticated guardrails with PreToolUse input modification
3. Skills hot reload eliminates restart friction for rapid development
4. Connect more MCP servers freely—lazy loading handles the complexity
5. This is the biggest Claude Code upgrade since launch (1,096 commits)

The constraint on MCP tool usage just disappeared. Build accordingly.

Have questions about implementing MCP Tool Search or hooks in your workflow? Check the AI Architecture Hub for blueprints and working examples.

Sources:

• Anthropic Claude Code Releases
• VentureBeat: Claude Code MCP Tool Search
• VentureBeat: Claude Code 2.1.0
• Claude Code Hooks Documentation
• Claude Code Skills Documentation

Related Articles

• Build Your Own Jarvis with Claude Code — The complete architecture guide
• What is Agentic AI? — Understanding the shift from tools to collaborators
• MCP Server Integration Guide — How Model Context Protocol works under the hood