TL;DR: Advanced Tool Use on Claude

Three features to cut tokens and boost accuracy - Tool Search, Programmatic Calling, and Usage Examples.

Claude's tool system eats context fast. 50 tools = 50K tokens gone before you ask anything.

              CONTEXT USAGE: TRADITIONAL vs TOOL SEARCH
                           (200k token context)

  Traditional approach (77.2k/200k tokens)
  ┌────────────────────────────────────────────────────────────┐
  │███████████████████████████████████████░░░░░░░░░░░░░░░░░░░░░│ 61.4% free
  └────────────────────────────────────────────────────────────┘
   ██ MCP tools: 72k  ▒ System: 5k  ░ Free: 122.8k

  Tool Search approach (8.7k/200k tokens)
  ┌────────────────────────────────────────────────────────────┐
  │████░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░│ 95.6% free
  └────────────────────────────────────────────────────────────┘
   ██ Search tool: 0.5k  ▓ Discovered: 3k  ▒ System: 5k  ░ Free: 191.3k

  Token savings: 68.5k tokens → 89% reduction in tool overhead

Three Features

1. Tool Search Tool

Mark tools with defer_loading: true - they won't load until Claude searches for them.

85% token reduction while keeping access to full tool library
Only search tool loads initially (~500 tokens)
Best for: 10+ tools, MCP-powered systems

Instead of each tool result returning to Claude, Claude writes a Python script that orchestrates the entire workflow. The script runs in the Code Execution tool (sandboxed), pausing when it needs results from your tools. Tool results are processed by the script rather than consumed by the model. Claude only sees the final output.

                  PROGRAMMATIC TOOL CALLING FLOW

  User          API         Code Execution       Claude
   │              │                │                │
   │──Request────►│                │                │
   │              │────Sampling───────────────────► │
   │              │                │   Claude generates Python script
   │              │◄────────────────────────────────│
   │              │                │                │
   │              │──Run script───►│                │
   │              │                │                │
   │              │◄─Pause: need───│                │
   │◄─Tool call───│   tool result  │                │
   │              │                │                │
   │──Tool result►│                │                │
   │              │──Resume───────►│                │
   │              │      ...       │  (loop for each tool)
   │              │◄─Script done───│                │
   │              │                │                │
   │              │────Sampling: script result─────►│
   │              │◄──────Claude interprets─────────│
   │◄──Response───│                │                │

37% token reduction on complex research tasks
Mark tools with allowed_callers: ["code_execution_20250825"]
Best for: Large datasets, multi-step workflows, parallel operations
20+ tool calls in ONE inference pass - intermediate results stay in sandbox

3. Tool Use Examples

Show Claude correct parameter patterns beyond JSON Schema.

{
  "tool_use_examples": [
    {
      "name": "create_event",
      "input": {
        "date": "2024-03-15",
        "attendees": ["user:123", "user:456"]
      }
    }
  ]
}

72% → 90% accuracy on complex parameter handling
Best for: Nested structures, domain conventions, many optional params

Accuracy Gains

Model	Before	After (Tool Search)
Opus 4	49%	74%
Opus 4.5	79.5%	88.1%

When to Use What

Problem	Solution
Context bloat (10K+ tokens in tools)	Tool Search
Many intermediate results	Programmatic Calling
Parameter formatting errors	Tool Use Examples

Reference

Advanced Tool Use on Claude Developer Platform - Anthropic Engineering