Toolofficial publisher

tool_search (Anthropic)

Tool : A precise action the AI can call · requires anthropic-api

Lets Claude find the right tool among hundreds without loading all schemas upfront.

The right reflex as soon as an agent has more than twenty tools.

Docs

Overall score

7.7 /10

Security: nothing to flag

Utility

7.0

Reliability

6.0

Security

No caveat on record

Ease

5 min · Technical

Security reflects where data goes, requested permissions and incidents on record at the verification date. It is not a guarantee of zero risk: check the deployment mode and the connections you enable.

What it does, concretely

When a Claude agent has hundreds of tools available, loading all their definitions into memory consumes a significant number of tokens before Claude even begins working. A typical setup combining GitHub, Slack, Sentry, Grafana, and Splunk can use around 55,000 tokens just for definitions, reducing selection accuracy and inflating every request.

Tool search solves this by letting Claude discover and load only the tools it needs, on demand, from a catalog of hundreds or thousands. Tools marked as deferred stay server-side; Claude issues a search query, the API returns up to five matching definitions, and Claude can then call the relevant tool. This approach reduces token consumption by more than 85% and keeps selection accuracy high even across thousands of tools.

This tool is aimed at developers building agents or automation pipelines with the Anthropic API who manage a tool catalog exceeding around twenty entries. It is available on Claude Haiku 4.5, Sonnet 4.5 and 4.6, Opus 4.5 through 5, Fable 5 and 5.1, and Mythos 5 and 5.1.

It is not suited to users who do not access the Anthropic API directly, nor to those whose tool catalog stays below around thirty entries, for whom direct loading remains simpler.

What it offers

·Regex-based search (regex variant): Claude constructs Python re.search() patterns to find matching tools, with case-insensitive matching.
·Natural language search (BM25 variant): Claude formulates a free-text query to search the tool catalog.
·Deferred tool loading: definitions marked with defer_loading: true are only injected into context when Claude discovers them through search.
·Up to 10,000 results per search: the limit parameter, ranging from 1 to 10,000, controls the maximum number of tools returned (default: 5).
·Prompt caching compatibility: deferred loading preserves the system-prompt prefix cache, as tool_reference blocks are appended inline without modifying the prefix.
·MCP integration: for tools from MCP servers via the MCP connector, deferred loading is configured at the mcp_toolset entry level rather than on each tool individually.
·Custom client-side implementation: the server-side search can be replaced with custom logic (embeddings, semantic search) by returning tool_reference blocks from a custom tool.

Our tips

Keep your 3 to 5 most frequently used tools without defer_loading, so Claude can call them directly without going through a search.
Never set defer_loading: true on the tool search tool itself, as this would make it inaccessible from the start.
For the regex variant, keep patterns under 200 characters and prefer simple expressions like "weather" or "get_.*_data"; for the BM25 variant, the query can be up to 500 characters of natural language.
Always send the full tools array (including deferred definitions) with every request: the API needs them server-side to run the search and resolve tool_reference blocks.
Never return a tool_result for the srvtoolu_... ID generated by tool search: the API will reject the request.

Watch out

The computer use and browser use toolsets (computer_toolset and browser_toolset) do not support defer_loading at the top-level entry; it must be set per member tool inside the configs object, and all enabled member values must match, otherwise the request is rejected.
A search returning no results produces an empty tool_references array with no explicit error message, which can make debugging less straightforward.

In practice

Price: Free
Commercial use: Commercial use: yes
Available on: api
Host: anthropic-api
FR interface: No

Security and data

Rating: Security: nothing to flag
Capabilities: readswritesexecutesnetwork
Your data: Data with the vendor
Maintenance: Active

Links and repository checked on 25 October 2026 · added on 1 January 2026

How to install tool_search (Anthropic)

1. Obtain an Anthropic API key from platform.claude.com and store it in an environment variable (never hard-code it). 2. Install the Anthropic SDK for your language (for example: pip install anthropic for Python). 3. In your messages.create call, add the search tool to the tools array with its type (tool_search_tool_regex_20251119 or tool_search_tool_bm25_20251119). 4. Add defer_loading: true to each tool in your catalog that Claude should not load upfront; never apply it to the tool search tool itself. 5. Send the full tools array with every request, even for deferred tools, so the API can run searches server-side. 6. Check model compatibility: Claude Opus 4.1 and earlier models do not support tool search.

pip install anthropic

Official documentation ↗

In the same category

Whole category Orchestrate →