Exa MCP Server

Exa MCP Server on Amazon Bedrock AgentCore Runtime

Exa is a search engine built for AI. This container runs Exa's open-source MCP server, giving agents neural web search and clean page extraction.

Get an Exa API key

Subscribe to the Exa Search API in AWS Marketplace to buy Exa credits on your AWS bill: https://aws.amazon.com/marketplace/pp/prodview-qwnuhblljqu72 

Your API key is available in the Exa dashboard: https://dashboard.exa.ai 

This container listing is free. You only pay for Exa API usage.

Prerequisites

• AWS CLI v2, updated to the latest version.
• An IAM role that Amazon Bedrock AgentCore can assume. Create a role with a custom trust policy using the JSON below, then attach permissions to pull the Marketplace container image and write CloudWatch logs.

{ "Version": "2012-10-17", "Statement": [ { "Effect": "Allow", "Principal": { "Service": "bedrock-agentcore.amazonaws.com" }, "Action": "sts:AssumeRole" } ] }

Create the agent runtime

aws bedrock-agentcore-control create-agent-runtime
--region us-east-1
--agent-runtime-name "<AGENT_NAME>"
--description "<AGENT_DESCRIPTION>"
--role-arn "<AGENT_ROLE_ARN>"
--agent-runtime-artifact '{"containerConfiguration":{"containerUri":"<IMAGE_URI_FROM_LAUNCH_PAGE>"}}'
--network-configuration '{"networkMode":"PUBLIC"}'
--protocol-configuration '{"serverProtocol":"MCP"}'
--environment-variables '{"EXA_API_KEY":"<YOUR_EXA_API_KEY>"}'

The response includes the agentRuntimeArn used to invoke the runtime.

Tools

Enabled by default:

• web_search_exa: Neural web search returning ranked results with page content.
• web_fetch_exa: Fetch and parse URLs as clean text.

Set ENABLED_TOOLS to enable advanced tools:

--environment-variables '{"EXA_API_KEY":"<YOUR_EXA_API_KEY>","ENABLED_TOOLS":"web_search_exa,web_fetch_exa,web_search_advanced_exa,agent_run"}'

• web_search_advanced_exa: Fine-grained control over search type, date and domain filters, and text, highlights, and summary content options.
• agent_run: Agentic research for multi-step tasks. Long-running requests return status "running" with a runId. Call the tool again with that runId to poll for completion.

Payloads

{"jsonrpc":"2.0","id":1,"method":"tools/list"}

{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"web_search_exa","arguments":{"query":"latest research on retrieval augmented generation","numResults":5}}}

{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"web_fetch_exa","arguments":{"urls":["https://exa.ai"],"maxCharacters":5000 }}}

web_fetch_exa takes urls as an array.

Write queries as descriptive statements of what you want. Exa matches meaning rather than keywords.

Container details

• Stateless streamable HTTP MCP server on 0.0.0.0:8000
• MCP endpoint: POST /mcp
• Health endpoint: GET /ping
• linux/arm64
• Runs as non-root user uid 10001

Support

Docs: https://docs.exa.ai  Source: https://github.com/exa-labs/exa-mcp-server  Email: support@exa.ai