Scale to many tools with tool search
Scale your agent to thousands of tools by discovering and loading only what's needed, on demand.
Tool search enables your agent 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, the agent searches your tool catalog and loads only the tools it needs.
This approach solves two challenges as tool libraries scale:
- Context efficiency: Tool definitions can consume large portions of the context window (50 tools can use 10-20K tokens), leaving less room for actual work.
- Tool selection accuracy: Tool selection accuracy degrades with more than 30-50 tools loaded at once.
Tool search is enabled by default.
How tool search works
When tool search is active, tool definitions are withheld from the context window. The agent receives a summary of available tools and searches for relevant ones when the task requires a capability not already loaded. Up to five of the most relevant tools are loaded into context by default, where they stay available for subsequent turns. If the conversation is long enough that the SDK compacts earlier messages to free space, previously discovered tools may be removed, and the agent searches again as needed.
Tool search adds one extra round-trip the first time Claude discovers a tool (the search step), but for large tool sets this is offset by smaller context on every turn. With fewer than ~10 tools, loading everything upfront is typically faster.
For details on the underlying API mechanism, see Tool search in the API.
Configure tool search
Tool search is on by default. It is disabled by default on Google Cloud's Agent Platform, where it is supported for Claude Sonnet 4.5 and later and Claude Opus 4.5 and later. It is also disabled when ANTHROPIC_BASE_URL points to a non-first-party host, since most proxies do not forward tool_reference blocks. You can override either default with the ENABLE_TOOL_SEARCH environment variable:
| Value | Behavior |
|---|---|
| (unset) | Tool search is on. Tool definitions are deferred and discovered on demand. Falls back to loading upfront on Google Cloud's Agent Platform or a non-first-party ANTHROPIC_BASE_URL. |
true |
Tool search is always on. The SDK sends the beta header even on Google Cloud's Agent Platform and through proxies. Requests fail on Google Cloud's Agent Platform models earlier than Sonnet 4.5 or Opus 4.5, or on proxies that do not support tool_reference blocks. |
auto |
Checks the combined token count of all tool definitions against the model's context window. If they exceed 10%, tool search activates. If they're under 10%, all tools are loaded into context normally. |
auto:N |
Same as auto with a custom percentage. auto:5 activates when tool definitions exceed 5% of the context window. Lower values activate sooner. |
false |
Tool search is off. All tool definitions are loaded into context on every turn. |
Setting CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS keeps tool search off, and ENABLE_TOOL_SEARCH can't override it. The variable strips the beta header that defer_loading tool definitions and tool_reference content blocks require.
Tool search applies to all registered tools, whether they come from remote MCP servers or custom SDK MCP servers. When using auto, the threshold is based on the combined size of all tool definitions across all servers.
Set the value in the env option on query(). In TypeScript, env replaces the subprocess environment, so spread ...process.env to keep inherited variables. In Python, env is merged on top of the inherited environment. This example connects to a remote MCP server that exposes many tools, pre-approves all of them with a wildcard, and uses auto:5 so tool search activates when their definitions exceed 5% of the context window:
import { query } from "@anthropic-ai/claude-agent-sdk";
try {
for await (const message of query({
prompt: "Find and run the appropriate database query",
options: {
mcpServers: {
"enterprise-tools": {
// Connect to a remote MCP server
type: "http",
url: "https://tools.example.com/mcp"
}
},
allowedTools: ["mcp__enterprise-tools__*"], // Wildcard pre-approves all tools from this server
env: {
...process.env, // env replaces the subprocess environment, so keep inherited variables
ENABLE_TOOL_SEARCH: "auto:5" // Activate tool search when tools exceed 5% of context
}
}
})) {
if (message.type === "result" && message.subtype === "success") {
console.log(message.result);
}
}
} catch (error) {
// A single-shot query() throws after yielding an error result
console.log(`Session ended with an error: ${error}`);
}
import asyncio
from claude_agent_sdk import query, ClaudeAgentOptions, ResultMessage
async def main():
options = ClaudeAgentOptions(
mcp_servers={
"enterprise-tools": {
"type": "http",
"url": "https://tools.example.com/mcp",
}
},
allowed_tools=[
"mcp__enterprise-tools__*"
], # Wildcard pre-approves all tools from this server
env={
"ENABLE_TOOL_SEARCH": "auto:5" # Activate tool search when tools exceed 5% of context
},
)
try:
async for message in query(
prompt="Find and run the appropriate database query",
options=options,
):
if isinstance(message, ResultMessage) and message.subtype == "success":
print(message.result)
except Exception as error:
# A single-shot query() raises after yielding an error result
print(f"Session ended with an error: {error}")
asyncio.run(main())
To run this example, replace https://tools.example.com/mcp with the URL of your own MCP server. On success the result text prints to the console.
Because this is a single-shot query() call, the SDK raises after yielding an error result, so the example wraps the loop in a try block. To see why a run failed, check the result message's subtype, such as error_during_execution, inside the loop. For more on result messages, see Handle the result.
Setting ENABLE_TOOL_SEARCH to "false" disables tool search and loads all tool definitions into context on every turn. This removes the search round-trip, which can be faster when the tool set is small (fewer than ~10 tools) and the definitions fit comfortably in the context window.
Optimize tool discovery
The search mechanism matches queries against tool names and descriptions. Names like search_slack_messages surface for a wider range of requests than query_slack. Descriptions with specific keywords ("Search Slack messages by keyword, channel, or date range") match more queries than generic ones ("Query Slack").
You can also add a system prompt section listing available tool categories. This gives the agent context about what kinds of tools are available to search for. Pass the text through the systemPrompt option in TypeScript or system_prompt in Python, using the claude_code preset with append, which adds your text to the preset's prompt instead of replacing it:
options: {
systemPrompt: {
type: "preset",
preset: "claude_code",
append: "You can search for tools to interact with Slack, GitHub, and Jira."
}
}
options = ClaudeAgentOptions(
system_prompt={
"type": "preset",
"preset": "claude_code",
"append": "You can search for tools to interact with Slack, GitHub, and Jira.",
}
)
For the full set of system prompt options, see Modifying system prompts.
Limits
- Maximum tools: 10,000 tools in your catalog
- Search results: returns up to five most relevant tools per search by default
- Model support: Claude Sonnet 4.5, Claude Haiku 4.5, Claude Opus 4.5, and later models; see model compatibility in the API docs for the current list. On Google Cloud's Agent Platform, Claude Sonnet 4.5 and later and Claude Opus 4.5 and later.
Related documentation
- Tool search in the API: Full API documentation for tool search, including custom implementations
- Connect MCP servers: Connect to external tools via MCP servers
- Custom tools: Build your own tools with SDK MCP servers
- TypeScript SDK reference: Full API reference
- Python SDK reference: Full API reference