Voozh

URL: https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool

Tool search

The tool search tool enables Claude to work with hundreds or thousands of tools by dynamically discovering and loading them on-demand. Instead of loading all tool definitions into the context window upfront, Claude searches your tool catalog (including tool names, descriptions, argument names, and argument descriptions) and loads only the tools it needs.

This approach solves two problems that compound quickly as tool libraries scale:

Context bloat: Tool definitions eat into your context budget fast. A typical multi-server setup (GitHub, Slack, Sentry, Grafana, Splunk) can consume ~55k tokens in definitions before Claude does any actual work. Tool search typically reduces this by over 85%, loading only the 3–5 tools Claude actually needs for a given request.
Tool selection accuracy: Claude's ability to correctly pick the right tool degrades significantly once you exceed 30–50 available tools. By surfacing a focused set of relevant tools on demand, tool search keeps selection accuracy high even across thousands of tools.



For background on the scaling challenges that tool search solves, see Advanced tool use. Tool search's on-demand loading is also an instance of the broader just-in-time retrieval principle described in Effective context engineering.

Although this is provided as a server-side tool, you can also implement your own client-side tool search functionality. See Custom tool search implementation for details.



Share feedback on this feature through the feedback form.



This feature is eligible for Zero Data Retention (ZDR). When your organization has a ZDR arrangement, data sent through this feature is not stored after the API response is returned.

Was this page helpful?



On Amazon Bedrock, server-side tool search is available only through the InvokeModel API, not the Converse API.



On Claude Platform on AWS, server-side tool search works identically to the Claude API. Claude Platform on AWS uses the Anthropic Messages API directly, so there is no InvokeModel or Converse distinction.

How tool search works

There are two tool search variants:

Regex (tool_search_tool_regex_20251119): Claude constructs regex patterns to search for tools
BM25 (tool_search_tool_bm25_20251119): Claude uses natural language queries to search for tools

When you enable the tool search tool:

You include a tool search tool (for example, tool_search_tool_regex_20251119 or tool_search_tool_bm25_20251119) in your tools list.
You provide all tool definitions with defer_loading: true for tools that shouldn't be loaded immediately.
Claude sees only the tool search tool and any non-deferred tools initially.
When Claude needs additional tools, it searches using a tool search tool.
The API returns 3-5 most relevant tool_reference blocks.
These references are automatically expanded into full tool definitions.
Claude selects from the discovered tools and calls them.

This keeps your context window efficient while maintaining high tool selection accuracy.

Quick start

Here's a simple example with deferred tools:

client = anthropic.Anthropic()

response = client.messages.create(
 model="claude-opus-4-8",
 max_tokens=2048,
 messages=[{"role": "user", "content": "What is the weather in San Francisco?"}],
 tools=[
 {"type": "tool_search_tool_regex_20251119", "name": "tool_search_tool_regex"},
 {
 "name": "get_weather",
 "description": "Get the weather at a specific location",
 "input_schema": {
 "type": "object",
 "properties": {
 "location": {"type": "string"},
 "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]},
 },
 "required": ["location"],
 },
 "defer_loading": True,
 },
 {
 "name": "search_files",
 "description": "Search through files in the workspace",
 "input_schema": {
 "type": "object",
 "properties": {
 "query": {"type": "string"},
 "file_types": {"type": "array", "items": {"type": "string"}},
 },
 "required": ["query"],
 },
 "defer_loading": True,
 },
 ],
)

print(response)

Tool definition

The tool search tool has two variants:

JSON

{
 "type": "tool_search_tool_regex_20251119",
 "name": "tool_search_tool_regex"
}

JSON

{
 "type": "tool_search_tool_bm25_20251119",
 "name": "tool_search_tool_bm25"
}



Regex variant query format: Python regex, NOT natural language

When using tool_search_tool_regex_20251119, Claude constructs regex patterns using Python's re.search() syntax, not natural language queries. Common patterns:

"weather" - matches tool names/descriptions containing "weather"
"get_.*_data" - matches tools like get_user_data, get_weather_data
"database.*query|query.*database" - OR patterns for flexibility
"(?i)slack" - case-insensitive search

Maximum query length: 200 characters



BM25 variant query format: Natural language

When using tool_search_tool_bm25_20251119, Claude uses natural language queries to search for tools.

Deferred tool loading

Mark tools for on-demand loading by adding defer_loading: true:

JSON

{
 "name": "get_weather",
 "description": "Get current weather for a location",
 "input_schema": {
 "type": "object",
 "properties": {
 "location": { "type": "string" },
 "unit": { "type": "string", "enum": ["celsius", "fahrenheit"] }
 },
 "required": ["location"]
 },
 "defer_loading": true
}

Key points:

Tools without defer_loading are loaded into context immediately
Tools with defer_loading: true are only loaded when Claude discovers them through search
The tool search tool itself should never have defer_loading: true
Keep your 3-5 most frequently used tools as non-deferred for optimal performance

Both tool search variants (regex and bm25) search tool names, descriptions, argument names, and argument descriptions.

How deferral works internally: Deferred tools are not included in the system-prompt prefix. When the model discovers a deferred tool through tool search, the API appends a tool_reference block inline in the conversation, then expands it into the full tool definition before passing it to Claude. The prefix is untouched, so prompt caching is preserved. The grammar for strict mode (the rules that constrain tool-call output to match your schemas) builds from the full toolset, so defer_loading and strict mode compose without grammar recompilation.

Response format

When Claude uses the tool search tool, the response includes new block types:

JSON

{
 "role": "assistant",
 "content": [
 {
 "type": "text",
 "text": "I'll search for tools to help with the weather information."
 },
 {
 "type": "server_tool_use",
 "id": "srvtoolu_01ABC123",
 "name": "tool_search_tool_regex",
 "input": {
 "query": "weather"
 }
 },
 {
 "type": "tool_search_tool_result",
 "tool_use_id": "srvtoolu_01ABC123",
 "content": {
 "type": "tool_search_tool_search_result",
 "tool_references": [{ "type": "tool_reference", "tool_name": "get_weather" }]
 }
 },
 {
 "type": "text",
 "text": "I found a weather tool. Let me get the weather for San Francisco."
 },
 {
 "type": "tool_use",
 "id": "toolu_01XYZ789",
 "name": "get_weather",
 "input": { "location": "San Francisco", "unit": "fahrenheit" }
 }
 ],
 "stop_reason": "tool_use"
}

Understanding the response

server_tool_use: Indicates Claude is calling the tool search tool
tool_search_tool_result: Contains the search results with a nested tool_search_tool_search_result object
tool_references: Array of tool_reference objects pointing to discovered tools
tool_use: Claude calling the discovered tool

The tool_reference blocks are automatically expanded into full tool definitions before being shown to Claude. You don't need to handle this expansion yourself. It happens automatically in the API as long as you provide all matching tool definitions in the tools parameter.

MCP integration

For configuring mcp_toolset with defer_loading, see MCP connector.

Custom tool search implementation

You can implement your own tool search logic (for example, using embeddings or semantic search) by returning tool_reference blocks from a custom tool. When Claude calls your custom search tool, return a standard tool_result with tool_reference blocks in the content array:

JSON

{
 "type": "tool_result",
 "tool_use_id": "toolu_your_tool_id",
 "content": [{ "type": "tool_reference", "tool_name": "discovered_tool_name" }]
}

Every tool referenced must have a corresponding tool definition in the top-level tools parameter with defer_loading: true. This approach lets you use more sophisticated search algorithms while maintaining compatibility with the tool search system.



The tool_search_tool_result format shown in the Response format section is the server-side format used internally by Anthropic's built-in tool search. For custom client-side implementations, always use the standard tool_result format with tool_reference content blocks as shown in the preceding example.

For a complete example using embeddings, see the tool search with embeddings cookbook.

Error handling



The tool search tool is not compatible with tool use examples. If you need to provide examples of tool usage, use standard tool calling without tool search.

HTTP errors (400 status)

These errors prevent the request from being processed:

All tools deferred:

{
 "type": "error",
 "error": {
 "type": "invalid_request_error",
 "message": "All tools have defer_loading set. At least one tool must be non-deferred."
 }
}

Missing tool definition:

{
 "type": "error",
 "error": {
 "type": "invalid_request_error",
 "message": "Tool reference 'unknown_tool' has no corresponding tool definition"
 }
}

Tool result errors (200 status)

Errors during tool execution return a 200 response with error information in the body:

JSON

{
 "type": "tool_search_tool_result",
 "tool_use_id": "srvtoolu_01ABC123",
 "content": {
 "type": "tool_search_tool_result_error",
 "error_code": "invalid_pattern"
 }
}

Error codes:

too_many_requests: Rate limit exceeded for tool search operations
invalid_pattern: Malformed regex pattern
pattern_too_long: Pattern exceeds 200 character limit
unavailable: Tool search service temporarily unavailable

Common mistakes

Prompt caching

For how defer_loading preserves prompt caching, see Tool use with prompt caching.

The system automatically expands tool_reference blocks throughout the entire conversation history, so Claude can reuse discovered tools in subsequent turns without re-searching.

Streaming

With streaming enabled, you'll receive tool search events as part of the stream:

event: content_block_start
data: {"type": "content_block_start", "index": 1, "content_block": {"type": "server_tool_use", "id": "srvtoolu_xyz789", "name": "tool_search_tool_regex"}}

// Search query streamed
event: content_block_delta
data: {"type": "content_block_delta", "index": 1, "delta": {"type": "input_json_delta", "partial_json": "{\"query\":\"weather\"}"}}

// Pause while search executes

// Search results streamed
event: content_block_start
data: {"type": "content_block_start", "index": 2, "content_block": {"type": "tool_search_tool_result", "tool_use_id": "srvtoolu_xyz789", "content": {"type": "tool_search_tool_search_result", "tool_references": [{"type": "tool_reference", "tool_name": "get_weather"}]}}}

// Claude continues with discovered tools

Batch requests

You can include the tool search tool in the Messages Batches API. Tool search operations through the Messages Batches API are priced the same as those in regular Messages API requests.

Limits and best practices

Limits

Maximum tools: 10,000 tools in your catalog
Search results: Returns 3-5 most relevant tools per search
Pattern length: Maximum 200 characters for regex patterns
Model support: Claude Fable 5, Claude Mythos 5, Claude Mythos Preview, Sonnet 4.0+, Opus 4.0+, Haiku 4.5+

When to use tool search

Good use cases:

10+ tools available in your system
Tool definitions consuming >10k tokens
Experiencing tool selection accuracy issues with large tool sets
Building MCP-powered systems with multiple servers (200+ tools)
Tool library growing over time

When traditional tool calling might be better:

Less than 10 tools total
All tools are frequently used in every request
Very small tool definitions (<100 tokens total)

Optimization tips

Keep 3-5 most frequently used tools as non-deferred
Write clear, descriptive tool names and descriptions
Use consistent namespacing in tool names: prefix by service or resource (for example, github_, slack_) so that search queries naturally surface the right tool group
Use semantic keywords in descriptions that match how users describe tasks
Add a system prompt section describing available tool categories: "You can search for tools to interact with Slack, GitHub, and Jira"
Monitor which tools Claude discovers to refine descriptions

Usage

Tool search tool usage is tracked in the response usage object:

JSON

{
 "usage": {
 "input_tokens": 1024,
 "output_tokens": 256,
 "server_tool_use": {
 "tool_search_requests": 2
 }
 }
}

Next steps

Tool reference

Full tool catalog with model compatibility and parameters.

MCP connector

Configure MCP toolsets with deferred loading.

Prompt caching

Combine tool search with cached tool definitions.

Define tools

Step-by-step guide for defining tools.

URL: https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool

⇱ Tool search tool - Claude API Docs

How tool search works

Quick start

Tool definition

Deferred tool loading

Response format

Understanding the response

MCP integration

Custom tool search implementation

Error handling

HTTP errors (400 status)

Tool result errors (200 status)

Common mistakes

Prompt caching

Streaming

Batch requests

Limits and best practices

Limits

When to use tool search

Optimization tips

Usage

Next steps