The built-in WebSearch is holding your agent back — switch to Exa

If your agent searches the internet, it's probably using the WebSearch that came bundled with Claude Code or Codex — and that default is quietly mediocre. It came free with the agent, so nobody questions it. You should: search is where your agent's research, lead-finding, and fact-checking all bottleneck.

Exa is a search engine purpose-built for agents, the API has a free tier, and wiring it into Claude Code takes about two minutes. This guide covers why the default falls short, what Exa actually is, the exact setup, and how to make Claude reach for it.

Agent search and human search are different jobs. You type one query and click a link. An agent doing one deep research task fans out into dozens of sub-searches, fires them in bursts, and consumes whole pages instead of snippets. The bundled WebSearch was built as a convenience feature, and it shows: generic results, consumer-search ranking, and every query filtered through infrastructure designed for people.

The tell is in your own transcripts. Watch Claude do research with the stock tool and you'll see it re-query, get listicles, and synthesize from SEO sludge. The quality ceiling of everything downstream — the report, the lead list, the code it writes against current docs — is set by what search returns.

Exa is a search engine built for exactly that agent traffic pattern. The scale numbers from CEO Will Bryk's recent post: Exa serves 80 billion pages, tracks 1.4 trillion URLs, and is on track for Google-scale coverage (~1 trillion pages) by early 2027 — against estimates of roughly 500B for Bing, 200B for Yandex, and 40B for Brave. For an independent index, that's a serious crawl.

The part I find more interesting than the size: they crawl far more than they serve, because most of the web is junk that actively degrades AI output — and they train models to filter it before it ever reaches your agent. Cleaner input, better synthesis. The pitch that sold me came from a developer who ran Perplexity, Exa, and Parallel side by side and settled on Exa for speed, cost, and result quality.

Grab a key at dashboard.exa.ai/api-keys — the free tier needs no card. Then register Exa's MCP server with Claude Code:

claude mcp add -e EXA_API_KEY=your-key-here exa -- npx -y exa-mcp-server

That runs the server locally over npx. If you'd rather not run anything, Exa also hosts it — point Claude at the remote endpoint instead:

claude mcp add --transport http exa "https://mcp.exa.ai/mcp?exaApiKey=your-key-here"

Restart Claude Code, then confirm it's connected:

claude mcp list

You'll see exa with a connected status, and Claude gains Exa tools for web search and page crawling alongside its built-ins.

One gotcha: installing the MCP doesn't remove the stock WebSearch — Claude now has both, and it won't always pick the better one. The fix is a single line in your project's CLAUDE.md (or the global one at ~/.claude/CLAUDE.md):

- For web research, prefer the Exa MCP tools over the built-in WebSearch.

Then prove the difference to yourself. Take a research question you actually care about and run it both ways:

Research <your question> twice: once using only the built-in WebSearch, once using only the Exa tools. Show me both answers with their sources, then compare — which sources are higher quality, which answer is more current, and which run needed fewer searches?

Watch for three things in the comparison: whether the sources are primary pages or roundup posts, how current the information is, and how many searches each run burned to get there. That's the whole pitch — more effective, cheaper, faster — and this one prompt lets you verify it on your own workload instead of taking anyone's word for it.

Defaults optimize for convenience, and this one costs you every research run. Search for agents is becoming its own infrastructure layer — Bryk predicts agents firing millions of queries per second within two years, and infrastructure built for that load and that consumption pattern will keep beating tools built for ten blue links. Your agent is only as smart as what it reads. Two minutes of setup upgrades the reading.