eve MCP connection: six HTTP requests per tool call

A minimal reproduction for vercel/eve: one turn that makes one MCP tool call sends eleven HTTP requests to the MCP server, of which one is the tool call.

mcp-server.mjs is both the server and the instrument — it implements a stateless Streamable-HTTP MCP server with one echo tool, and prints one line per request it receives.

Run it

npm install
echo 'AI_GATEWAY_API_KEY=<your key>' > .env.local   # or VERCEL_OIDC_TOKEN=…

node mcp-server.mjs                                 # terminal 1
npx eve invoke "Echo the word marigold."            # terminal 2

What terminal 1 prints

  1  GET /mcp  → 405            ← transport probe
  2  POST /mcp  initialize      ┐ connection_search's client
  3  POST /mcp  notifications/initialized
  4  GET /mcp  → 405            │
  5  POST /mcp  tools/list      ┘
  6  GET /mcp  → 405            ┐ the tool call's client — a new one
  7  POST /mcp  initialize      │
  8  POST /mcp  notifications/initialized
  9  GET /mcp  → 405            │
 10  POST /mcp  tools/list      │
 11  POST /mcp  tools/call      ┘

What is expected

One tool call should cost one tools/call, plus a handshake amortised over the session — not a fresh initialize + notifications/initialized + tools/list per workflow step, and not a GET probe before every one of them when the server has already answered POST.

Recorded on eve 0.44.3, Node 24, Linux. The same eleven-request shape was first seen on 0.34.0 against a production MCP server, where each request costs a serverless round trip rather than a loopback one.