Cheatsheet: Server-side tools and built-ins

Three tool categories (decision frame)

Category	Who executes	Schema you author	Examples
Server tools	Anthropic	Just type identifier	web_search, web_fetch, code_execution
Anthropic-schema client tools	You	Anthropic publishes schema	bash, computer use, memory, text_editor
Tool infrastructure	Anthropic	Mixed	tool_search, MCP connector (L6)

Decision: who owns the capability + who runs the code?

Server tools at a glance

Tool	Type identifier	Pricing	ZDR
web_search	`web_search_20260209` (dynamic filtering) / `web_search_20250305` (basic)	$10 / 1,000 searches ($0.01 each)	Basic: eligible. Dynamic filtering: NOT eligible
web_fetch	`web_fetch_20260209` (dynamic filtering) / `web_fetch_20250910` (basic)	Standard tokens (no per-fetch charge)	Basic: eligible. Dynamic filtering: NOT eligible
code_execution	`code_execution_20260120` / `code_execution_20250825`	FREE when used with `web_search_20260209` or `web_fetch_20260209`; otherwise standard per-execution pricing	NOT eligible

The free-when-used-with pricing rule

The canonical case: search plus computation with one tool declared. Dynamic filtering runs code internally:

tools=[{"type": "web_search_20260209", "name": "web_search"}]

If the task also needs top-level code execution beyond filtering search results (plotting, file writes, a calculation independent of search), declare both. The per-execution charge stays waived because web_search_20260209 is in the request:

tools=[
    {"type": "web_search_20260209", "name": "web_search"},
    {"type": "code_execution_20260120", "name": "code_execution"},
]

Computation only with no search or fetch: declare code_execution_20260120 alone and pay standard per-execution pricing.

web_search optional params

Param	Use for
`max_uses`	Cap searches per request (cost guardrail)
`allowed_domains`	Whitelist sources
`blocked_domains`	Blacklist sources
`user_location`	Localize results (approximate; city/region/country/timezone)

Citations: always on; web_search_result_location blocks with url / title / encrypted_index / up to 150 chars of cited_text. Display to end users per docs.

Anthropic-schema client tools

Tool	What it does
`bash`	Execute shell commands
`text_editor`	Create + edit files
`memory`	Persist knowledge across conversations
`computer use` (beta)	Screenshots + mouse + keyboard

Same L4 client loop (tool_use → execute → tool_result). Brief tool descriptions OK because the model already knows the canonical tool.

Computer use discipline (the careful section)

Beta header required: computer-use-2025-11-24 for Opus 4.8 / Opus 4.7 / Opus 4.6 / Sonnet 4.6 / Opus 4.5; earlier header for older models.
ZDR eligible. State-of-the-art on WebArena.
Sandbox the environment: VM, container, or isolated profile with minimum credentials. Never against real desktop or production-reachable systems. Treat the agent like a contractor with access only to what they need.
Track 20 (“AI Agents and Tool Use”) goes deeper on agent-design at track-level depth.

tool_search

tools = [
    {"type": "tool_search_tool_regex_20251119", "name": "tool_search_tool_regex"},
    {"name": "get_weather", "description": "...", "input_schema": {...}, "defer_loading": True},
    {"name": "search_files", "description": "...", "input_schema": {...}, "defer_loading": True},
    # ...many more deferred tools
]

Variant	Query type
`tool_search_tool_regex_20251119`	Python `re.search` patterns (max 200 chars)
`tool_search_tool_bm25_20251119`	Natural-language queries

	Use when	Skip when
tool_search	10+ tools / >10k definition tokens / selection accuracy issues / MCP with 200+ tools	<10 tools / all-used-frequently / <100-token total

Max 10,000 tools in catalog. Returns 3-5 per search.
tool_search tool itself NEVER has defer_loading: true.
Keep 3-5 most-used tools non-deferred.
Preserves prompt caching (deferred tools not in system-prompt prefix).

Server response shape

{
  "role": "assistant",
  "content": [
    { "type": "text", "text": "I'll search." },
    { "type": "server_tool_use", "id": "srvtoolu_...", "name": "web_search",
      "input": { "query": "..." } },
    { "type": "web_search_tool_result", "tool_use_id": "srvtoolu_...",
      "content": [ /* search results */ ] },
    { "type": "text", "text": "Based on...", "citations": [...] }
  ],
  "stop_reason": "end_turn"
}

server_tool_use + the matching tool-specific result block (web_search_tool_result, code_execution_tool_result, web_fetch_tool_result, tool_search_tool_result) INLINE.
No round-trip in your code.
Iterate response.content, surface text + citations.
stop_reason typically end_turn. Mid-loop: pause_turn (send back as assistant turn, re-call to continue).

Pricing implications

Total cost per call = standard tokens
                    + per-tool fees
                    + server-tool results in context (as input tokens on next turns)

Dials to manage:
  • Declare web_search_20260209 (or web_fetch_20260209) → code-execution per-execution charge is waived
  • max_uses on web_search → cap runaway
  • Cache stable tool definitions (L7) → drop repeated tool overhead
  • tool_search for large catalogs → keep most defs out of context
  • Eval-set discipline (L3 + Track 21 L7) → measure tool earns its cost

ZDR eligibility (check per-tool)

Tool	ZDR
web_search (basic: `web_search_20250305`)	Eligible
web_search (dynamic filtering: `web_search_20260209`)	NOT eligible
web_fetch (basic: `web_fetch_20250910`)	Eligible
web_fetch (dynamic filtering: `web_fetch_20260209`)	NOT eligible
code_execution	NOT eligible
computer use	Eligible
bash / text_editor / memory	Standard messages-API ZDR
tool_search	Eligible

Common pitfalls

Failure	Recognize by	Fix
Building your own web search	Wasted effort on solved problem	Drop in web_search server tool
Paying standalone code_execution when search or fetch is also in flight	Surprise per-execution charges	Declare web_search_20260209 or web_fetch_20260209 in the same request (per-execution charge is waived)
Computer use against real env	Real-world side effects	Sandbox: VM / container / isolated profile
tool_search below threshold	Complexity without benefit	Standard tool calling for <10 tools
Skipping citations	UX miss + docs-recommended discipline	Always display web_search citations to end users
Ignoring per-tool ZDR	Compliance gap	Check the table above before declaring

What this lesson does NOT cover (and where to find it)

Topic	Lands at
Model Context Protocol (cross-provider tools)	Lesson 6
Prompt caching on tool definitions	Lesson 7
The agent loop (model + tools across many turns; pause_turn in prod)	Lesson 8 onward
Subagents and Claude Managed Agents	Lesson 11
Advisor server tool (research preview)	Beyond T22; see docs

Source

Anthropic public Claude docs: Web search, Code execution, Web fetch, Computer use, Tool search, Server tools at https://platform.claude.com/docs/en/agents-and-tools/tool-use/. Structural-mirror citation; see references.