Exa runs a hosted MCP server that gives an agent live web search and clean page reads. It is unusually easy to try: the endpoint answers an initialize handshake without any credential at all, so you can connect a client and start searching in seconds.
https://mcp.exa.ai/mcp
Claude Haiku 4.5
Chat with 60+ AI models on the same workflow — switch to a different model mid-conversation and re-run the same prompt, or use Compare mode to put several side-by-side and balance quality vs. cost.
No token required
How models use it and what it is built for.
Exa's server is deliberately small — two tools rather than the sprawling surfaces common elsewhere — and that focus is the point. `web_search_exa` searches the live web and returns clean, ready-to-use content rather than a list of links a model then has to fetch and parse. `web_fetch_exa` reads any URL and returns it as clean markdown, which is the natural follow-up when a search highlight is not enough or when you already know the page you want. The pairing maps neatly onto how an agent actually researches: search broadly, then read deeply on the two or three results that matter. Because the server reports itself as exa-search-server with Exa's own metadata and website, there is no ambiguity about provenance.
Typical tools an AI model can call. Exact names vary by version.
Taken from the official Exa documentation — see Exa MCP documentation for the full reference.
Examples
Hosted endpoint
https://mcp.exa.ai/mcpStreamable HTTP. Responds to initialize with no credential, which makes it a genuinely good server to test a client against.
Client configuration
Any MCP client
No credential block needed — just the URL.
{
"mcpServers": {
"exa": {
"type": "http",
"url": "https://mcp.exa.ai/mcp"
}
}
}Copy any of these into MCP Agent Studio after connecting.
Search for what changed in the MCP specification this year and summarise the breaking changes.
Find three recent independent benchmarks of AI tool calling and compare their methodology.
Read this documentation page and tell me the exact auth flow it describes.
Who are the main competitors in this space right now, and what does each claim to do differently?
This is not a single-model product: you get the same MCP connection with 60+ models (Claude, GPT, Gemini, DeepSeek, open-weight, and more), you can switch mid-conversation, and you can open Compare mode to run the same prompt against multiple models at once. The card above is a suggested starting point for this server — not the only choice.
Default pick for Exa
Claude Haiku 4.5
Only two tools, so selection is trivial and the work is summarising what comes back. Haiku handles that at a fraction of the cost of a frontier model.
Listing tools proves the server is reachable, not that a model can work with it. Evals go further: they read every tool on the server, write a test suite from its real schemas, and run it — code decides pass/fail on the responses (schema conformance, error codes, pagination, result caps) while a scoring model grades plain-English tasks driven through the tools.
Get a pass/fail report per tool with the evidence behind each verdict — and replay the same suite after every schema change. Destructive tools are excluded from the run.
Run evalsOpen MCP Agent Studio with the connection pre-filled. Add your token, pick any of 60+ models, and start chatting — no install required.
Open Agent StudioCommon questions about connecting, scoping and using it safely.
Not to get started. The hosted endpoint at https://mcp.exa.ai/mcp completes an MCP initialize handshake with no credential at all, which is rare — almost every hosted server answers 401 first. That makes it one of the easiest real servers to point a new client at. An Exa API key is the route to higher limits once you are past experimenting.
Two: `web_search_exa`, which searches the live web and returns clean content rather than bare links, and `web_fetch_exa`, which reads any URL as clean markdown. That is the whole surface. The restraint is deliberate and it helps — with two tools, every model picks correctly, which is not true of servers exposing hundreds.
All three give an agent web access and all three are worth testing on your own workload. The distinction that matters in practice is what comes back: Exa returns cleaned, ready-to-use content and pairs search with a dedicated markdown fetch tool, which suits research workflows where the model needs to actually read sources. Run the same prompt against each in Agent Studio — result quality varies more by query type than any comparison table can capture.
Yes, and it is worth knowing about for that reason alone. Because it needs no auth, it removes the credential variable when you are debugging a client — if your client cannot connect to Exa, the problem is your client, not your token. For deliberately awkward cases, pair it with our mock servers, which return errors and nested schemas on demand.
It can read them. `web_fetch_exa` exists precisely for the follow-up step: search returns highlights, and when those are insufficient the model fetches the full page as markdown. This two-step pattern is why the server works well for research tasks rather than only fact-lookup, and it is worth writing prompts that expect it.
Brave Search
Give an AI agent live web and local search via the Brave Search API.
Tavily
Search, extract and crawl the web for AI research agents via Tavily.
Perplexity
Get cited, web-grounded answers from an AI agent via Perplexity Sonar.
Wikipedia
Ground AI answers in verifiable Wikipedia articles — no API key needed.