Key-free web tools. Source: extensions/web.ts.
- WebSearch: DuckDuckGo search with
allowed_domains/blocked_domains, no API key. - WebFetch: SSRF-guarded fetch that prefers markdown via
Acceptand converts HTML otherwise. - WebFetch reads up to 10 MiB and converts the whole page before capping the output at 30,000 characters, so content that sits behind hundreds of KB of markup (a GitHub pull request's conversation) is reached. A larger response is refused with an error rather than cut, since a cut inside a
<script>or<style>leaks its contents as text. - An optional
promptruns the page through the model in-process and returns the answer, falling back to markdown when headless or on error.
- http is upgraded to https before the request, as Claude documents. The upgraded URL is also the cache key, so both spellings of a page share one entry.
- A redirect to a different host is reported, not followed: the result names the original URL and the target, and Claude fetches the target with a second call if it wants it. Same-host redirects are still followed, up to 5 of them, with the address guard re-applied per hop. Because a cross-host target is never requested, it cannot be reached from here at all.
- Cache TTL is 15 minutes per URL by default and honors
CLAUDE_CODE_WEBFETCH_CACHE_TTL_MS. An unset, blank, negative or non-numeric value keeps the default. - A page is decoded as the charset its
content-typedeclares, not always as UTF-8 (the Fetch spec's own.text()hardcodes UTF-8, which returned mojibake for Windows-1252 and ISO-8859-1 pages). A header naming no charset, or oneTextDecoderdoes not know, falls back to UTF-8 rather than failing the fetch. - Esc cancels a fetch or search in flight, not just its nested
promptcompletion: the tool call's own signal is combined with the 20s per-hop timeout, so either can end the request, and a redirect chain is checked between hops too.
The cross-host rule is WebFetch's. The internal search fetch still follows its own endpoint's redirects: those hops are ours, and Claude's rule does not describe them.
- User-Agent is pi-code's own, not
Claude-User*. The tools reference does not document the string, so there is nothing to match. - WebSearch
allowed_domains+blocked_domainsin one call: pi-code applies allowed and ignores blocked. Claude's tool description calls them mutually exclusive; the reference gives no behavior for the combination. - No 200-call per-session WebSearch cap. Not in the reference either; recorded here because Claude's product surface has one.
- Page extraction runs on the session model, where Claude uses a small fast model.