Exa Plugs 100B Websites Directly Into OpenAI's Codex Agent

OpenAI's coding agent just got a serious upgrade to its research skills. Exa, the AI-native search company, has released an official plugin that lets Codex and ChatGPT tap directly into its index of the live web, papers, docs, companies, and people. Installation is a three-click affair from the Codex plugin directory.

The plugin lands inside OpenAI's newly unified plugin ecosystem. OpenAI recently migrated its App directory into a Plugin directory, making plugins the primary way to discover workflow capabilities across ChatGPT and Codex. Both products share one universal plugin directory, so the same public plugins are discoverable from their supported surfaces, working in Chat and Work across ChatGPT web, desktop, and mobile, and in Codex in the ChatGPT desktop app. The Codex CLI has its own plugin browser as well.

What the plugin actually does

Once installed, Codex can call Exa mid-task to pull live web context instead of relying on stale training data or generic browsing. That matters for the kind of work coding agents get stuck on: hunting down the latest API changes, cross-referencing library docs, or grounding a decision in a real paper. According to the plugin registry, the integration lets Codex search and fetch current web sources through the official Exa MCP server with browser OAuth.

The pitch, straight from Exa's launch, is that Codex now gets a pipe into more than 100 billion websites, docs, papers, profiles, and company records. Under the hood, that's backed by infrastructure Exa has been building for years:

• Custom vector databases designed for extremely high throughput
• Sub-200 millisecond latency for real-time agents, plus a higher-quality tier for asynchronous agents that need comprehensive data
• Embedding and agentic models trained for retrieval, with vector DBs built from scratch for 10k+ QPS
• Zero data retention on queries, since Exa built its search engine from scratch and offers ZDR for all product endpoints

Why now

Exa is riding the wave of agents doing more searching than humans. In May, the company announced a $250M Series C, with founder Will Bryk writing that AI agents will search the web more than humans this year, and Exa raised at a $2.2B valuation led by a16z to power all agents with the highest quality web search.

The company already sits underneath a good chunk of the AI stack. Exa powers search for Cursor, Cognition, HubSpot, OpenRouter, Monday.com and over 400,000 developers, and expects agent searches to grow 1000x beyond Google searches in the coming years. Landing directly in Codex extends that reach into what is arguably the highest-value agent surface in the market right now.

The context that matters

OpenAI has been pushing Codex out of its niche as a standalone coding tool. The company has been pushing Codex beyond an agentic coding desktop app to a general productivity tool, and is putting Codex functionality inside the ChatGPT app everywhere. Plugins are the connective tissue for that expansion, and web search is the most obvious gap in a coding agent that otherwise lives inside your repo.

Exa's positioning here is worth understanding. The company has spent years arguing that most search APIs sold to AI companies are thin wrappers. As Bryk puts it, there are more space programs than true search engines, which is why most companies selling AI search are in fact wrappers over existing search engines. Owning the crawl, the embeddings, and the vector store is expensive, but it's what lets Exa offer things generic search wrappers cannot, like true zero-retention queries and custom classifiers over results.

What it unlocks in practice

A few workflows become materially easier with Exa wired into Codex:

1. Grounding code generation in current library docs rather than whatever the model memorized during pretraining
2. Pulling structured lists from Websets, Exa's high-compute product that produces massive lists of people, companies or really anything you want, straight into an agent loop
3. Cross-referencing research papers when working on ML code or algorithms
4. Doing competitive or vendor research from inside a coding session without breaking flow to open a browser

For anyone already paying for a ChatGPT plan that includes Codex, the plugin is free to install from the directory. The bigger question is whether Exa's grounding actually improves task success rates for real coding work, particularly against OpenAI's own built-in browsing tools. Given that six months ago Exa was worse than Google at code search and is now used by nearly every coding agent, the honest answer is that it will vary by task, and it is worth benchmarking on your own workload before making it the default.