Exa is a web search API built for AI applications, and exa-py is its official Python client, targeting Python 3.9 and later. One client object covers four surfaces: search over the web index, content retrieval for URLs, a direct answer endpoint, and an agent API for longer research runs. An AsyncExa variant mirrors the interface for asynchronous code.

search() returns ranked results and can request highlights or full text with them. You can cap the result count, bound the publication date, restrict to named domains, pass a systemprompt to steer selection, and supply an outputschema that returns either plain text or structured JSON — object schemas are limited to a nesting depth of 2 and 10 total properties. streamsearch() yields OpenAI-style chat completion chunks. Deep search variants deep-lite, deep, and deep-reasoning also accept additional queries. getcontents() fetches text, highlights, or summaries for URLs you already hold, and answer() answers a question directly, with a streaming form.

The client also emits Exa as a tool definition for model loops. exa.openai.websearch() and exa.anthropic.websearch() produce ready-made tool specifications, exa.openai.responses.websearch() covers the OpenAI Responses API, and handletoolcalls executes the calls and returns their outputs. Tools can be renamed, which lets the Exa search tool run alongside a provider's own built-in search tool where names must be unique. The handlers answer every tool call, returning an explicit unknown-tool error rather than dropping one, so a follow-up request stays valid; when you run other tools alongside Exa's, you replace those error outputs with your own results before the next request. getcontents lives in the same namespaces, so a model can read pages whose URLs it already has.

The Agent API takes a query, an output schema, and an effort level, starts a run, and is polled until it finishes; a beta namespace adds a maximum effort level with a cost budget. Agent Monitors, also in beta, keep a table of entities by fields fresh on a cadence: static fields are answered once per entity from the live web, and dynamic fields are tracked from news at every refresh. You can page a monitor's entities, follow a cursor-resumable change feed, add entities, inspect refresh progress, and delete the monitor, or take a one-shot stateless snapshot over a past news window without creating one.

Features

Search: ranked web results with optional highlights or full text, domain filters, date bounds, and a result-count cap
Structured output: a systemprompt and a text or JSON outputschema on every search type, with streaming through stream_search()
Deep search: deep-lite, deep, and deep-reasoning variants that accept additional queries
Contents: text, highlights, and summaries for supplied URLs, with a live-crawl preference option
Answer: a direct answer endpoint with a streaming variant
Model-loop tools: prebuilt websearch and getcontents tool definitions for OpenAI chat completions, the OpenAI Responses API, and Anthropic messages
Tool-call handling: a handler that runs the calls and answers every one, including unresolvable tool names
Agent API: schema-constrained agent runs with effort levels, polling, and an optional cost budget
Agent Monitors: cadence-refreshed entity and field tables with a change feed and stateless snapshots
Async client: an AsyncExa counterpart for asynchronous applications

Claim Exa

Verify write access with GitHub