Monetize the Context Layer:
The MCP Server Blueprint

-- detect_mcp_tool_call()
-- Runs in log phase — AFTER response sent.
-- Zero latency added to inference hot path.

local function detect_mcp_tool_call(body)
  local ok, parsed = pcall(cjson.decode, body)
  if not ok then return nil end

  if parsed.jsonrpc ~= "2.0" then return nil end
  if parsed.method ~= "tools/call" then return nil end

  local tool_name = parsed.params.name
  local agent_id  = kong.request
                      .get_header("X-Agent-Id")

  -- Buffer in lua_shared_dict, flush async
  -- to POST /v1/ingest/batch
  return {
    tool_name  = tool_name,
    agent_id   = agent_id,
    tenant_id  = kong.client.get_consumer().custom_id,
    started_at = ngx.now() * 1000,
  }
end

-- Wired in log phase:
function AforoMeteringHandler:log()
  local event = detect_mcp_tool_call(body)
  if event then
    aforo_buffer:push(event)   -- non-blocking
  end
end

import { AforoClient } from "@aforo/mcp-metering";

const billing = new AforoClient({
  tenantId:  process.env.AFORO_TENANT_ID,
  productId: process.env.AFORO_PRODUCT_ID,
  apiKey:    process.env.AFORO_API_KEY,
});

// One-line decorator. Works for stdio and SSE.
// Measures wall-clock duration automatically.
const wrapped = billing.wrapToolHandler(
  async (name: string, args: unknown) => {
    // Record execution duration + token metadata
    // Flush batch → POST /v1/ingest/batch
    // Heartbeat ACK carries kill signal back
    return await yourToolRouter(name, args);
  }
);

// Register as your MCP tool dispatcher:
server.setRequestHandler(
  CallToolRequestSchema,
  wrapped
);

{
  "metricName":
    "mcp_server.tool_invocations",
  "productType": "MCP_SERVER",

  "toolName":  "web_search",
  "sessionId": "sess_abc123",

  "agentId":
    "agent_gpt4o_prod",
  "executionDurationMs": 542,

  "metadata": {
    "inputTokens":  1200,
    "outputTokens": 450
  }
}