Web research & live sources

1 · Three tools, one posture

• web_search — a concise, bounded search of the public web.
• open_web_page — a single bounded page read as text.
• web_research — one search plus up to three source-page reads in a single bounded action; results come back focus-ranked with page excerpts.

All three are read-only, public-HTTPS only, and work out of the box — no MCP connector, no API key, no terminal permission required. Access levels are controlled in Settings → Tools & Access like every other tool.

2 · The guardrails

• Local and private hosts are blocked — no poking at your router, your NAS, or localhost services from the agent.
• Non-text downloads are blocked — the research tools don't pull binaries onto your disk.
• Oversized pages are refused — bounded reads, not infinite scraping.
• Verified-domain search filters let you restrict results to domains you trust, with focus-ranked excerpts so you actually read the good part.

3 · What it's good at

Docs lookups ("what does this flag do in the latest release?"), release notes, error-message archaeology, dependency research, and comparing sources before you commit to an approach. Activity streams into the chat as collapsed blocks and is persisted with the session, so you can audit exactly which pages fed an answer.

4 · What it deliberately isn't

It is not a headless browser, not a crawler, and not an archive. It won't log you into anything, execute page scripts as your identity, or bypass site rules. For interactive sites, use the Plugin Hub's Sites workflow or paste the relevant content into chat.

5 · Privacy posture

Search and page reads make normal HTTPS requests to public sites — those servers see a request from your machine, because that's how the web works. Your prompts, workspace and credentials are not attached. No research content is stored anywhere except your local session (Security & privacy →).