exa-code: fast, efficient web context for coding agents (exa.ai)

Exa released exa-code, a web-scale context tool built specifically for coding agents that returns the exact few hundred tokens from the web needed to ground code generation. Rather than dumping long web pages, exa-code prioritizes extremely dense, highly relevant snippets—especially code examples—so agents get compact, actionable context (typical responses are <500 tokens). You trigger it by adding "use exa-code" in the prompt. The goal: reduce brittle, library- and API-related hallucinations that still plague LLMs in 2025 by giving agents the precise examples and parameters they need. Technically, exa-code combines Exa’s AI-first search engine with a dedicated code-example index sourced from GitHub and Exa’s web index, plus code-specific retrieval models to maximize recall of relevant snippets. In their benchmark, tasks were generated from documentation and scored by an LLM classifier that detects hallucinated code; exa-code achieved the largest reduction in hallucinations while using far fewer tokens than other context sources. Implications include lower latency and cost for agent calls, fewer incorrect API calls, and improved reliability across popular and obscure dependencies—an important step toward agents that can reliably write and integrate real-world software.