Session ID is no longer being re-used by same Claude Desktop sessions

[naji247]

Describe the bug
We've noticed a regression across all Claude Desktop users where every tool call is causing a creating and using a new session ID despite them being within the same working session of the user or seconds apart.

To Reproduce
Steps to reproduce the behavior:
Create a simple Todo Server with durable workers/CF agents sdk.
Connect via Claude Desktop
@modelcontextprotocol/sdk: 1.24.0
@agentS: 0.2.24

Ask any questions to modify todo list, and every tool call has a different advertised session ID

Expected behavior
A clear and concise description of what you expected to happen.
Session ID is reused properly and Request ID goes up incrementally.

Screenshots
If applicable, add screenshots to help explain your problem.

Version:
@agentS: 0.2.24
@modelcontextprotocol/sdk: 1.24.0

[threepointone]

can you try with 0.4.0? tagging @mattzcarey

[threepointone]

can you try with 0.4.0? tagging @mattzcarey

[naji247]

@threepointone I can try suggesting it to our clients, let me report back. For convincing them to upgrade, is there a known issue in compatibility resolved?

[threepointone]

clients don't need to upgrade, I meant the server?

[naji247]

Sorry, we are working on this issue on behalf of a company (our client) haha. That is what I meant.

[mattzcarey]

Could I ask how you are building the MCP Servers? Are you using createMcpHandler or McpAgent and which Transport are you using (if applicable)

[naji247]

We're noticing this on McpAgent and it's happening to both Streamable and SSE sessions.

[threepointone]

@mattzcarey dod you think this has to do with them moving away from the query param thing? ( I can't recollect the details)

[mattzcarey]

I'll look. It might be :(

[naji247]

I think this is happening where User Agent is "Claude-User", which is the user-agent for "Claude Connector" extensions (not MCP servers custom added). We also noticed this started happening around Feb 8th.

[mattzcarey]

I have tested our McpAgent again and I cannot replicate using standard payloads from the MCP spec. Feels like a Claude Desktop issue.

[naji247]

Yes I would agree here. Sadly we don't have a way to contact them about this. Oh well. Shall I close this ticket?

[naji247]

FWIW I think it must be due to some of these clients not quite aware of the durable worker style deployments. Seems like they don't bother to use session IDs properly. I am noticing this with these user agents as well:

@n8n/n8n-nodes-langchain.mcpClient
Manus
mcporter

I'll follow up with them individually since their teams are more responsive 😓