Let Claude retrieve and analyze full page or PDF content with the server-side web fetch tool, with domain allowlists, citations, and content caps

Documented steps

1. POST /v1/messages. No anthropic-beta header is required — web fetch is generally available; stale guides that add a web-fetch beta header are out of date.
2. Add to tools: {"type":"web_fetch_20250910","name":"web_fetch","max_uses":5}. Newer: web_fetch_20260209 (dynamic filtering), web_fetch_20260309 (adds use_cache), web_fetch_20260318 (adds response_inclusion).
3. Restrict reachable hosts with allowed_domains OR blocked_domains (mutually exclusive).
4. Turn on citations explicitly with citations:{"enabled":true} — unlike web search, web fetch citations are OFF by default.
5. Cap injected content with max_content_tokens (e.g. 100000). This is approximate and applies to text content, not binary PDFs.
6. On web_fetch_20260309+, set use_cache:false to bypass Anthropic's result cache (default true) when freshness matters.
7. On web_fetch_20260318+, set response_inclusion:"excluded" to omit nested server_tool_use / result blocks.
8. Only supply URLs that the user provided explicitly or that appeared in a prior web_search / web_fetch result — server-side SSRF protection blocks Claude from fetching URLs it constructed itself.
9. Parse {"type":"web_fetch_tool_result","tool_use_id":...,"content":{"type":"web_fetch_result","url":...,"content":{"type":"document","source":{"type":"text"|"base64","media_type":"text/plain"|"application/pdf","data":...},"title":...},"retrieved_at":<ISO 8601>}}.
10. Official docs: https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-fetch-tool