⏳ In Brief
- Web Fetch lets Claude pull full web pages and PDFs, now in beta.
- Enable with the anthropic-beta header value web-fetch-2025-09-10.
- Works on Opus 4.1, Opus 4, Sonnet 4, and Haiku 3.5.
- Tight security model, only fetch URLs already present in the conversation.
- Optional citations, domain allow/deny lists, token caps, and clear error codes.
Breaking: Anthropic’s Web Fetch Brings Direct Retrieval To Claude Agents
Anthropic’s Web Fetch tool retrieves full content from specified web pages and PDFs, giving agents immediate access to source material for analysis. The feature ships as a beta capability for Claude developers.
Enable Web Fetch by sending the anthropic-beta header set to web-fetch-2025-09-10, then add the web_fetch_20250910 tool with your chosen limits. Supported models include Opus 4.1, Opus 4, Sonnet 4, and Haiku 3.5.
New on our API: web fetch.
Add the web fetch tool to your requests and Claude will fetch and analyze content from any webpage URL—no additional infrastructure needed. pic.twitter.com/g5ZFjK6Zvf
— Claude (@claudeai) September 10, 2025
What It Does, In Practice
When included, Claude decides whether to fetch based on your prompt and the available URLs, then analyses returned content with optional citations. PDF text extraction happens automatically.
Developers can surface clickable citations in their UI, so readers can verify claims against exact passages, improving auditability for enterprise workflows.
“The web fetch tool allows Claude to retrieve full content from specified web pages and PDF documents.” — Anthropic.
Security And Governance, URL Rules And Domain Controls
Claude cannot construct arbitrary links. It may only fetch URLs that already appear in the conversation, including prior search or fetch results, which reduces exfiltration risk.
Admins can restrict scope with allowed_domains, blocked_domains, and max_uses, and should consider Unicode homograph risks when defining filters for production.
“For security reasons, the web fetch tool can only fetch URLs that have previously appeared in the conversation context.” — Anthropic.
Anthropic recommends using trusted environments or non-sensitive data when enabling Web Fetch, given potential exfiltration concerns in mixed-trust deployments.
Setup And Parameters, Headers And Limits
Add the web_fetch tool with optional max_uses, allowed_domains, blocked_domains, citations.enabled, and max_content_tokens for context control.
Use the anthropic-beta header value web-fetch-2025-09-10 in all requests that include this tool, alongside your chosen model and token budget.
Key Parameters At A Glance
- max_uses limits total fetches per request
- allowed_domains or blocked_domains filters scope
- citations.enabled toggles inline evidence
- max_content_tokens caps included content
Errors, Caching, And Content Types
Web Fetch supports text and PDF only. The API returns structured errors such as invalid_input, url_not_accessible, and max_uses_exceeded. Maximum URL length is 250 characters.
Results may be served from a cache, so returned content can lag a live page update. The cache is managed automatically and may evolve by content type.
“Enable citations for fetched content” by setting "citations": {"enabled": true}, with “clickable” inline references required in end-user interfaces. — Anthropic.
Search Plus Fetch: How They Work Together
Web search can discover sources, then Web Fetch retrieves full text for deeper analysis. Claude chooses results, fetches details, then responds with cited passages when enabled.
This pattern supports research, compliance reviews, and summarisation flows where grounding and reproducibility matter as much as speed.
“Use web search to find relevant articles… Use web fetch to retrieve full content… Provide detailed analysis with citations.” — Anthropic.
Conclusion
Web Fetch adds controlled retrieval to Claude agents, with model support, clear limits, and optional citations for verifiable outputs. The design emphasises security, governance, and traceability.
Teams can now combine search discovery with full-text fetch, then render clickable evidence in their apps. That pairing boosts trust while keeping deployment risks in check.
📈 Latest AI News
10th September 2025
For the recent AI News, visit our site.
