exa MCP

The Exa MCP server connects an agent to Exa's search engine. By default it gives two tools: web_search_exa for searching the web and web_fetch_exa for reading pages as Markdown. You connect to the hosted endpoint https://mcp.exa.ai/mcp, anonymously with rate limits or signed in for more. The server is MIT licensed, and Exa bills for usage. We checked it on 31 October 2026.

Quick facts

What it does

Exa describes itself as a search engine built for AI. The MCP server is its way into agent tools. The README's one-line summary is: connect AI agents to Exa for web search, content fetching and multi-step research.

The two default tools cover most agent needs.

That pair is the usual loop. The agent searches, picks results, then fetches the pages worth reading. Because both return cleaned text, the agent spends fewer tokens than it would on raw HTML.

Optional tools

Two more tools are off by default.

You turn them on by adding a tools parameter to the URL. The README warns that this replaces the defaults, so you must list every tool you want:

https://mcp.exa.ai/mcp?tools=web_search_advanced_exa
https://mcp.exa.ai/mcp?tools=web_search_exa,web_fetch_exa,agent_run

Exa Agent needs authentication, either OAuth or an API key. The simple tools do not.

Fewer tools is a quiet advantage. Every tool definition sits in the agent's context, so a server with two tools costs little. Compare that with servers that register dozens. The Firecrawl MCP server, for example, registers 27 in its full profile.

Pricing

The MCP server is free to install. Exa bills for what the agent uses. From Exa's pricing page:

The page lists per-product rates:

Two readings of this. First, the free allowance is real but small. At the lowest search rate, $10 covers a few thousand requests, which is plenty for trying the server and not much for a heavy agent loop. Second, the Agent product is priced per run across a wide range, so set expectations before you let an agent call it freely.

Exa's MCP docs say search and fetch work with no API key at a free rate limit, and that signing in or adding a key moves you off it. They do not say which row of the table applies to web_search_exa, so treat the table as a guide to the API, not a quote for the MCP tool.

Install it

The hosted endpoint is the route the README leads with.

https://mcp.exa.ai/mcp

For anonymous use you add nothing else. To sign in, most clients prompt you with an OAuth flow. To force the login flow, for a shared connector for example, use https://mcp.exa.ai/mcp?login or https://mcp.exa.ai/mcp/oauth.

If you prefer an API key, get one from dashboard.exa.ai/api-keys and send it one of three ways: as ?exaApiKey= on the URL, as an Authorization: Bearer header, or as an x-api-key header. We would use a header. A key in a URL ends up in logs, shell history and config files that get committed.

Install it in each agent

The README documents Claude (a plugin), Codex, Cursor, VS Code and a list of other clients. Our adaptation is marked where the README only says "add manually". For how each agent names fields, see MCP clients.

Claude Code

The README gives the plugin route:

claude plugin install exa@claude-plugins-official

By hand, our adaptation using Claude Code's HTTP transport:

claude mcp add --transport http exa https://mcp.exa.ai/mcp

Codex

From the README:

codex mcp add exa --url https://mcp.exa.ai/mcp

The README also points to a ChatGPT plugin listing for Codex and ChatGPT.

Cursor

The README has a one-click install link for Cursor. By hand, add this to .cursor/mcp.json:

{
  "mcpServers": {
    "exa": { "url": "https://mcp.exa.ai/mcp" }
  }
}

Antigravity

The README lists Google Antigravity as "add manually to mcp_config.json". The path is ~/.gemini/config/mcp_config.json, and Antigravity uses serverUrl for Streamable HTTP. Our adaptation:

{
  "mcpServers": {
    "exa": { "serverUrl": "https://mcp.exa.ai/mcp" }
  }
}

OpenCode

The README says "add manually to opencode.json". Our adaptation, using OpenCode's remote type:

{
  "mcp": {
    "exa": {
      "type": "remote",
      "url": "https://mcp.exa.ai/mcp"
    }
  }
}

OpenCode's remote entries also take headers and an oauth setting, which is where an API key header or the OAuth flow would go.

Gemini CLI, VS Code and others

Gemini CLI is now for Code Assist enterprise customers only, and the README says to add Exa to ~/.gemini/settings.json there. VS Code has an install link in the README. The README also lists Windsurf, Zed, Warp, LM Studio, Replit, Kiro, Grok Build and v0, each with its own location.

Agent skills in the repo

The repository doubles as an Agent Plugin and ships two skills in skills/.

The README says MCP-only setups still get the tools, and skills add orchestration on top. If you use Claude Code, the plugin install brings the skills. For how skills load in each agent, see Agent skills.

Example prompts and what to expect

These are the kinds of requests the tools suit. We have not run them, so treat the outcomes as expectations from the tool descriptions.

• "Find the latest changelog for this library and tell me what changed in the last minor version." The agent searches, then fetches the changelog page as Markdown.
• "Search for three comparison articles on these two tools and summarise where they disagree." Search returns clean content, and fetch pulls the full text where needed.
• "Build a list of ten companies in this niche with a website and a one-line description." This is a job for agent_run, which is built for list building and structured output, and it needs sign-in.
• "Check what this docs page says about rate limits." Fetch alone is enough.

Exa's pricing page also describes coding-agent uses such as searching docs, repos and changelogs. That is a good fit for a coding agent, where up-to-date package notes matter.

Choosing between search and fetch

A small habit saves credits. Ask the agent to search first and read the result titles and excerpts before it fetches anything. Search returns content that is often enough to answer a factual question, and a fetch is only needed when the answer sits deeper in a page. The pricing page bills contents separately, at $1 per 1,000 pages per content type, so fetching ten pages for every search quietly multiplies the cost. A line in your project instructions such as "search first, fetch only the two best results" is cheap and effective.

It also helps to name the source type in your prompt. "Check the official docs" and "find recent discussion" lead the agent to different queries, and with a general search engine the wording of the question matters more than any setting.

Security notes

1. Anonymous use is rate limited, not private. Your queries go to Exa. Do not send private code, customer data or credentials in a search.
2. Keys in URLs leak. The URL parameter route puts the key in whatever logs the URL. Prefer OAuth or a header.
3. A hosted server sees your queries. Exa's Enterprise tier lists zero data retention. For other tiers, read the privacy terms before sending sensitive topics.
4. Results are untrusted input. Search results and fetched pages are text written by strangers. A page can contain instructions aimed at an agent. Keep write tools and shell access out of the same session where you can, or require confirmation.
5. Spend caps matter. A loop that calls search a thousand times costs money on a paid plan. Set budget alerts in the Exa dashboard.
6. Plugin installs run more than tools. The Claude plugin also installs skills. Read them as you would any prompt that steers your agent.

Limits and known issues

• Small default tool set. Two tools is simple, but if you want filters and dates you must enable the advanced tool through the URL.
• The `tools` parameter replaces defaults. Easy to miss, and a common reason the basic search tool "disappears".
• Pre-1.0 style versioning is not used. The npm version was 3.4.2, with no GitHub release tags, so track npm if you pin a local copy.
• Agent runs vary in cost. The page gives a range from $0.012 to $1.00 per run.
• Open issues. The repo showed 7 open issues on 31 October 2026, which is low. Check the tracker for any that affect your client.
• Hosted dependency. If the endpoint is down, so is the tool.

Alternatives

The MCP servers hub lists more.

Which agents it suits

• Claude Code and Codex have a one-line route.
• Cursor has a one-click link and a short JSON entry.
• Antigravity and OpenCode take a short URL entry.
• Research and coding agents gain most from search plus fetch.
• Teams that want one tool, not many, like the two-tool default.

Our take

We did not run the Exa server for this page. This is from the README, the pricing page and repo data.

What we like is the restraint. Two default tools, a hosted URL and no install step make it among the easiest servers to add. The option to widen the tool set through the URL is a sensible way to keep context small by default.

What to watch is the money and the data. The free allowance is generous for trying and thin for production, and a hosted search server sees every query your agent sends. Use OAuth, set a spending alert and keep sensitive topics out of searches.

For general research and coding lookups, it is a good first search server to try. If you need crawling and extraction at scale, compare it with Firecrawl.

A short setup checklist

1. Add the hosted URL to your agent.
2. Ask for a search, then a fetch of one result.
3. Sign in with OAuth when you hit rate limits.
4. Enable agent_run only if you need it, and list all tools in the URL.
5. Check your dashboard credit use after the first session.
6. Set a spending alert before enabling pay as you go.
7. Keep private material out of search queries.

Frequently asked questions

What does the Exa MCP server do?

It gives an agent web search and page fetch, with optional advanced search and multi-step agent runs, through a hosted endpoint.

Do I need an API key?

Not for basic use. Anonymous use is rate limited, and OAuth or a key raises limits and unlocks Exa Agent.

Is it free?

The server is MIT licensed. Exa's Starter plan includes $10 in monthly credits and a $10 onboarding bonus.

How do I turn on the optional tools?

Add a tools parameter to the URL and list every tool you want, because it replaces the defaults.

Can I run it locally?

An npm package exists, and the README leads with the hosted endpoint. Check Exa's docs for local steps.