Prevent MCP transport reuse by creating a server per request

[mattt]

Resolves #43

[Copilot]

Pull request overview

This PR addresses the MCP endpoint outage (#43) by avoiding reuse of MCP server connection state across requests, which can lead to transport/connect failures and HTTP 500s.

Changes:

• Removed the in-memory MCP server cache keyed by external-doc allowlist/blocklist.
• Updated /mcp to instantiate a fresh MCP server for each incoming request before connecting the HTTP transport.

Comments suppressed due to low confidence (1)

src/index.ts:76

• Creating a new MCP server on every /mcp request re-registers all tools and Zod schemas each time, which adds per-request CPU/latency overhead and can increase cost under load. If the goal is only to avoid reusing a connected transport, consider keeping a cached, pre-initialized tool registry (or another lightweight factory) so you don’t rebuild the entire server for every request, while still ensuring each transport gets its own isolated connection state.

  const mcpServer = createMcpServer({
    EXTERNAL_DOC_HOST_ALLOWLIST: c.env.EXTERNAL_DOC_HOST_ALLOWLIST,
    EXTERNAL_DOC_HOST_BLOCKLIST: c.env.EXTERNAL_DOC_HOST_BLOCKLIST,
  })
  const transport = new StreamableHTTPTransport()
  await mcpServer.connect(transport)
  return transport.handleRequest(c)

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.