Tavily

Tavily is a web access API for AI agents. It retrieves live web data, extracts the relevant content, and returns it structured and chunked for a model, so an agent can ground its answers in current sources. This entry covers the Python SDK, which wraps the API's search, extract, crawl, map, research, and feedback endpoints in a synchronous TavilyClient and an AsyncTavilyClient.

Search takes a query and returns ranked results, with options for exact-phrase matching, a ready-made context string for retrieval-augmented generation (RAG), or a short direct answer. Extract pulls the raw content and images from up to 20 URLs per request. Crawl traverses a site from a base URL under natural-language instructions (invite-only), and Map returns a site's structure the same way. Research creates a multi-source report task with citations, which you poll for a result or stream as it runs.

Feedback lets an agent or end user score a search request and its individual results, which Tavily uses to tune ranking. Requests pass through security and content validation layers that block personally identifiable information (PII), prompt injection, and malicious sources.

A keyless mode runs search() and extract() without an API key at a rate-limited tier, and the other endpoints need a key. For teams that route traffic through an API gateway, the client accepts a preconfigured HTTP session in place of an API key. Optional project, session, user, and client identifiers attribute requests, and the user identifier is hashed server-side. A free tier covers personal use up to 1,000 credits per month.