Tool search tool

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

Loading every tool definition up front causes two problems as a tool library grows:

• Context bloat: A typical multi-server setup (GitHub, Slack, Sentry, Grafana, and Splunk) can consume ~55k tokens in definitions before Claude does any work. Tool search typically reduces this by over 85 percent, loading only the 3–5 tools Claude needs for a given request.
• Tool selection accuracy: Claude's ability to pick the right tool degrades once you exceed 30–50 available tools. Because tool search loads only a focused set of relevant tools on demand, selection accuracy stays high even across thousands of tools.

Tool search is generally available on the Claude API. For supported models, see Model compatibility.

Tool search runs as a server-side tool, but you can also implement your own client-side tool search. See Custom tool search implementation for details.



Model compatibility

Both tool search variants are available on the following models:

Claude Opus 4.1 and earlier models don't support the tool search tool.



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:

1. You include a tool search tool (for example, tool_search_tool_regex_20251119 or tool_search_tool_bm25_20251119) in your tools list.
2. You provide every tool definition in the tools array and set defer_loading: true on the tools that shouldn't load up front. At least one tool, normally the tool search tool itself, must stay non-deferred.
3. Initially, Claude's context contains only the tool search tool and any non-deferred tools.
4. When Claude needs additional tools, it searches using a tool search tool.
5. The API runs the search and returns the matching tools as tool_reference blocks (up to 5 by default).
6. The API automatically expands these references into full tool definitions.
7. Claude selects from the discovered tools and calls them.



Quick start

The following example includes the tool search tool and two 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)

Claude searches the catalog, discovers get_weather, and calls it. The response ends with stop_reason: "tool_use". Execute the discovered tool and return a tool_result as in Handle tool calls. Response format shows the blocks you get back and what to send next.



Tool definition

The tool search tool has two variants:

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

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



Deferred tool loading

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

{
  "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
}

defer_loading controls what enters the context window, not what you send in the request:

• You still send every tool's full definition in the tools array on every request, including the deferred ones. The API needs them server-side to run the search and expand tool_reference blocks.
• Tools without defer_loading load into context immediately.
• Tools with defer_loading: true load only when Claude discovers them through search.
• Never set defer_loading: true on the tool search tool itself.
• Keep your 3–5 most frequently used tools non-deferred so Claude can call them without searching first.

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

Internally, the API excludes deferred tools from the system-prompt prefix. When Claude 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 the following block types:

{
  "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": {
        "pattern": "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: Claude's call to the tool search tool. The search runs on Anthropic's servers. Never return a tool_result for its srvtoolu_... ID.
• tool_search_tool_result: the search results, in a nested tool_search_tool_search_result object. Keep it in the message history as is.
• tool_references: an array of tool_reference objects pointing to discovered tools. The API expands these for Claude. You never expand them yourself.
• tool_use: Claude's call to a discovered tool. Execute it and return a tool_result exactly as in standard tool use.

The API automatically expands tool_reference blocks into full tool definitions before showing them to Claude. You don't need to handle this expansion yourself, as long as you provide all matching tool definitions in the tools parameter.



Continuing the conversation

On the next request, pass the assistant's content back unchanged, including the server_tool_use and tool_search_tool_result blocks. Add your tool_result for the discovered tool in a user message, and send the same tools array: the search tool plus every deferred definition. Don't return a tool_result for the srvtoolu_... ID: the API rejects the request. The API expands tool_reference blocks throughout the conversation history, so Claude can reuse discovered tools in later turns without re-searching. A search that matches nothing returns a tool_search_tool_search_result with an empty tool_references array, not an error.



MCP integration

If your tools come from MCP servers through the MCP connector, you don't set defer_loading on individual tool definitions. Instead, set it once on the mcp_toolset entry's default_config for the whole server, or per tool in its configs. See MCP toolset configuration.



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:

{
  "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, normally with defer_loading: true. This lets you use search methods the built-in variants don't provide, such as embedding-based retrieval, and the API expands the returned tool_reference blocks the same way.

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



Error handling



HTTP errors (400 status)

These errors prevent the API from processing the request:

All tools deferred:

{
  "type": "error",
  "error": {
    "type": "invalid_request_error",
    "message": "At least one tool must have defer_loading=false. All tools cannot be deferred."
  }
}

Missing tool definition:

{
  "type": "error",
  "error": {
    "type": "invalid_request_error",
    "message": "Tool reference 'unknown_tool' not found in available tools"
  }
}



Tool result errors (200 status)

When a tool search operation fails during execution, the API returns a 200 response with the error in the body:

{
  "type": "tool_search_tool_result",
  "tool_use_id": "srvtoolu_01ABC123",
  "content": {
    "type": "tool_search_tool_result_error",
    "error_code": "invalid_tool_input",
    "error_message": "Invalid regular expression pattern: missing ) at position 1"
  }
}

The error_code field has four possible values:

• invalid_tool_input: the search input was invalid, for example a malformed regex pattern or a pattern over the 200-character limit
• unavailable: the search couldn't run, for example because it timed out or the service was unavailable
• too_many_requests: rate limit exceeded for tool search operations
• execution_time_exceeded: the search exceeded its execution time limit



Common mistakes



Prompt caching

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

A tool with defer_loading: true can't also carry cache_control: the API returns a 400. Put the cache breakpoint on a non-deferred tool.



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 pattern streamed
event: content_block_delta
data: {"type": "content_block_delta", "index": 1, "delta": {"type": "input_json_delta", "partial_json": "{\"pattern\":\"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.



Limits and best practices



Limits

• Maximum deferred tools: 10,000 tools with defer_loading: true per request
• Search results: each search returns up to 5 matching tools by default
• Pattern and query length: maximum 200 characters for regex patterns and 500 characters for BM25 queries
• Model support: see Model compatibility



When to use tool search

Use tool search when any of the following apply:

• You have 10 or more tools available.
• Your tool definitions consume more than 10k tokens.
• Tool selection accuracy drops as your toolset grows.
• You aggregate multiple MCP servers (200+ tools).
• Your tool library grows over time.

Standard tool calling, without tool search, is a better fit when you have fewer than 10 tools, every tool is used in every request, or your tool definitions are small (less than 100 tokens total).



Optimization tips

• Keep your 3–5 most frequently used tools 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 one search matches the whole group.
• Use 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 your descriptions.



Usage

Tool search isn't metered as a separate server tool. The response's usage.server_tool_use object has no tool search field, and the tool definitions that search loads into context count as input tokens like any other tool definition.



Next steps