McpAgent crashes with 'Already connected to a transport' after Durable Object hibernation

[ebrainte]

Bug Description

McpAgent crashes with "Already connected to a transport" when the Durable Object wakes up from hibernation. This happens because onStart() calls server.connect(transport) but the MCP Server instance still holds a reference to the previous transport from before hibernation.

Steps to Reproduce

1. Deploy any worker using McpAgent (e.g., the cloudflare/playwright-mcp example)
2. Connect an MCP client via SSE or streamable-http
3. Wait for the Durable Object to hibernate (idle timeout)
4. Send another request — the DO wakes up, onStart() runs, and crashes

Error

Error: Already connected to a transport
    at McpAgent.onStart (agents/dist/mcp/index.js)

Root Cause

In packages/agents/src/mcp/index.ts, the onStart() method (and the SSE/streamable-http handlers) call server.connect(transport) without first disconnecting the previous transport. After hibernation, the Server instance is restored but still thinks it's connected.

Three locations need the fix:

• onStart() (~line 183)
• SSE handler in handleMcpMessage() (~line 229)
• Streamable-HTTP handler in handleMcpMessage() (~line 240)

Fix

Add server.close() before each server.connect() call:

const server = await this.server;
try { await server.close(); } catch (_e) { /* not connected yet on first run */ }
await server.connect(this._transport);

The try/catch is needed because on the very first connection, server.close() may throw since no transport was ever connected.

Fix Branch

A complete fix is available at: ebrainte/agents:fix/mcp-agent-hibernation-reconnect

I was unable to create a PR because this repository restricts pull requests to collaborators only.

Affected Version

agents@0.0.109 (and current main as of March 2026)

Workaround

Use patch-package to apply the fix locally. Patch file for agents@0.0.109:

--- a/node_modules/agents/dist/mcp/index.js
+++ b/node_modules/agents/dist/mcp/index.js
@@ -170,6 +170,7 @@ var McpAgent = class _McpAgent extends DurableObject {
     );
     await this._init(this.props);
     const server = await this.server;
+    try { await server.close(); } catch (_e) { /* not connected yet */ }
     if (this._transportType === "sse") {
       this._transport = new McpSSETransport(() => this.getWebSocket());
       await server.connect(this._transport);
@@ -228,6 +229,7 @@ var McpAgent = class _McpAgent extends DurableObject {
         this._transportType = "sse";
         if (!this._transport) {
           this._transport = new McpSSETransport(() => this.getWebSocket());
+          try { await server.close(); } catch (_e) {}
           await server.connect(this._transport);
         }
         return this._agent.fetch(request);
@@ -238,6 +240,7 @@ var McpAgent = class _McpAgent extends DurableObject {
             (id) => this.getWebSocketForResponseID(id),
             (id) => this._requestIdToConnectionId.delete(id)
           );
+          try { await server.close(); } catch (_e) {}
           await server.connect(this._transport);
         }
         await this.ctx.storage.put("transportType", "streamable-http");

[threepointone]

This isn't because of hibernation, it's because you're reusing a transport instance (which the mcp sdk started blocking in 1.26.0 because of a cve GHSA-345p-7cg4-v4c7). I recommend that you update all versions of agents/ mcp sdk, and make sure you're not reusing server/transport instances across requests.

I'm closing this, but happy to reopen if you have a repro.

[ebrainte]

Thanks @threepointone — you're absolutely right. We upgraded to agents@^0.4.0 (which pulls in @modelcontextprotocol/sdk@1.26.0+) and removed all agent patches. The Durable Object hibernation wake-up now works correctly with fresh Server/transport instances per connection.

We've also updated our upstream PR (cloudflare/playwright-mcp#54) to use the version upgrade approach instead of the server.close() patch. Appreciate the quick response and the CVE context — that was exactly the insight we needed.