Source: Server tools (Claude docs)

Anthropic's shared mechanics for server-executed tools: server_tool_use blocks, the server-side loop and pause_turn, mixing server and client tools, allowed_callers and domain filtering.

Contents
  • What: official documentation page
  • Publisher: Anthropic
  • Date: living document; read 2026-11-14
  • Original: Server tools

Summary

Explains how tools that Anthropic executes (web search, web fetch, code execution, tool search) appear in a Messages response and how to continue turns that pause or mix in your own tools.

Key claims

  • server_tool_use blocks need no tool_result from you; the result block follows in the same turn, paired by tool_use_id.
  • Server tools run in a server-side loop; a long turn can return stop_reason: "pause_turn". Continue by resending the assistant content unchanged with the same tools (and the container ID for code execution); a turn can pause repeatedly, so cap continuations.
  • If Claude calls a server tool and a client tool in one parallel group, the API returns stop_reason: "tool_use" without running the server tool; send a user message containing only tool_result blocks and the server tool runs at the start of the next request.
  • Adding text after the tool results ends the turn and causes a 400 for the unresolved server call.
  • allowed_callers chooses direct calls, calls from code execution, or both; dynamic-filtering versions are not ZDR-eligible unless set to ["direct"].
  • Domain filters: no scheme, subdomains included, path wildcards only, allow and block lists mutually exclusive, request lists must be a subset of org-level lists; beware Unicode homographs.
  • In batches the loop has a higher per-turn iteration limit.

Feeds

Anthropic Messages API

Suggest changes →