A JavaScript MCP server deployed to AWS Bedrock AgentCore Runtime. Exposes tools via the Model Context Protocol that can be consumed by any MCP client (Cursor, Claude Desktop, Amazon Q CLI, etc.).

It serves the same canonical MCP server as the aws-mcp-servers example family - tools (add, slow_report with streamed progress, approve_refund with an elicitation round-trip), resources (guide://usage plus an orders://{orderId} template), a prompt, instructions served via server/discover, and cache hints - so the family's shared test client works here too, and you can compare hosting on AgentCore Runtime against the Lambda-based options side by side.

• Node.js 24.x
• AWS account with credentials configured
• Serverless Framework installed (npm i -g serverless)
• Docker running (the runtime is deployed as a container)

mcp-server/
├── index.js          # HTTP layer: node:http + toNodeHandler on port 8000
├── src/server.mjs    # the canonical MCP server (official MCP SDK v2)
├── client.mjs        # the family's shared test client
├── package.json      # Dependencies and start script
├── serverless.yml    # Serverless Framework configuration
└── README.md

npm install
sls deploy

The deploy output prints the runtime's MCP endpoint URL. The shared client calls it directly - the same way real MCP clients connect - signing requests with your local AWS credentials (SigV4, the runtime's default inbound auth):

AUTH=sigv4 SERVICE=bedrock-agentcore ENDPOINT=<agent URL from deploy output> node client.mjs

PASS  1. tools/list returns the canonical tools with cache hints — ttlMs=300000 cacheScope=public
PASS  2. add returns plain JSON with structured content — sum=42
PASS  3. slow_report streams incremental progress over SSE — 3 progress events, first at +810ms
PASS  4. approve_refund elicitation round-trip (accept and cancel) — input_required -> refunded o-1 / refund cancelled
...

Note the requestHeaders.allowlist block in serverless.yml: protocol revision 2026-07-28 requires the Mcp-Method and Mcp-Name headers on every request, and the allowlist forwards them to the container.

npm install
sls dev

The server is built with the official MCP TypeScript SDK v2: createMcpHandler produces a web-standard handler serving the stateless MCP 2026-07-28 protocol revision (self-contained requests, server/discover, structured results) while answering older MCP clients through the SDK's built-in fallback on the same endpoint. toNodeHandler mounts it on a plain node:http server at /mcp on port 8000, the standard expected by AgentCore Runtime for MCP-protocol runtimes. The Runtime adds an Mcp-Session-Id header for its own session isolation; the stateless server accepts and ignores it.

The serverless.yml sets protocol: MCP on the agent, which tells AgentCore to route MCP traffic to the runtime. A commented authorizer.jwt block shows platform OAuth: the runtime validates Bearer JWTs from any OpenID Connect provider before requests reach the container - all configuration, no code.

Related example: mcp-server-lambda-tools inverts this trade — AWS runs the MCP server (AgentCore Gateway) and your code shrinks to plain Lambda functions with the tool schema declared in serverless.yml. Pick this example when you need full SDK-level control; pick that one for simple request/response tools with no MCP code at all.

Once deployed, the runtime can be connected to any MCP client that supports remote MCP servers via Streamable HTTP. Refer to the AgentCore MCP documentation for invocation details and authentication setup.