Tavily
Search the web, fetch pages and crawl a site through the shared websearch family.
Current capabilities
The tavily adapter implements websearch.search, websearch.fetch and websearch.crawl as
profile tavily/2026-10, mapped from Tavily's pinned OpenAPI document. Every operation is a read
that Tavily serves as a POST; the host admits it through post_json for the three fixed paths
only, with no write approval.
Results carry each website's URL, title where Tavily gives one, description, content and
published date, with complete, truncation and provenance. A content longer than 128 KiB is
clipped, and items past a 3 MiB result are omitted. Search and crawl have answered live through
the local CLI; fetch has been exercised against a local fake server only.
Access and limits
The credential profile tavily.api-key holds the API key in Secret Service custody; it is
checked with GET /usage, which spends no search credit. Tavily exposes no account identifier,
so two keys are not told apart. Every read spends Tavily credits, and nothing is cached.
The repository's local Tavily guide covers the native configuration, connecting with the key on stdin, invoking each operation and replacing the executable.