anthropics / anthropics/claude-ai-mcp
Cowork/Code "sibling probe" for local MCP servers fails silently for servers that don't support concurrent instances
- Lingua principale
- Nessun dato sulla lingua
- Stelle
- 471
- Fork
- 76
- Metriche di merge delle PR
- Nessuna PR unita negli ultimi 30g
Descrizione
### What happened?
Claude Desktop logs show it attempts to launch a second ("sibling") instance of the server to probe its protocol era/capabilities. This sibling fails to complete the handshake:
[Macuse] [info] Era probe verdict: legacy (sibling did not complete the exchange)
In the main chat session this is silently tolerated (falls back to "legacy" mode, server works fine). But when Cowork/Code tries to spin up its own independent instance of the same server, the identical failure is fatal:
[Macuse] [error] Couldn't start this server for Cowork and Code sessions (they run their own copy of it), so they can't use its tools: Version negotiation failed: the connection closed during the server/discover probe (this transport probed in place — the disposable sibling probe requires the SDK's base StdioClientTransport) { metadata: { context: 'shared-pool', stack: undefined } }
First observed 2026-08-18; the server (Macuse) worked in this exact config since June 2026 with no errors before that date, suggesting this "sibling probe" mechanism is new.
### What did you expect to happen?
I expected my locally-configured MCP server (Macuse) to be available in Cowork and Code sessions, the same way it works in normal Claude Desktop chat sessions.
### Steps to reproduce
1. Configure a local stdio MCP server in Claude Desktop (in my case: Macuse, a macOS automation server) that does not support being launched as two concurrent instances.
2. Confirm the server works normally in a regular Claude Desktop chat session.
3. Open a Cowork or Code session.
4. Attempt to use a tool from that MCP server.
5. See error in Claude Desktop's MCP debug logs (View Logs for that server).
### Area
MCP Connector (adding/managing servers)
### MCP Server (if applicable)
_No response_
### Error messages or logs
```shell
[Macuse] [info] Era probe verdict: legacy (sibling did not complete the exchange)
[Macuse] [error] Couldn't start this server for Cowork and Code sessions (they run their own copy of it), so they can't use its tools: Version negotiation failed: the connection closed during the server/discover probe (this transport probed in place — the disposable sibling probe requires the SDK's base StdioClientTransport) { metadata: { context: 'shared-pool', stack: undefined } }
```
### Additional context
This appears to be a new mechanism ("sibling probe" / "Era probe") that launches a second, parallel instance of a locally-configured stdio MCP server — used both to detect its protocol era and to give Cowork/Code sessions their own independent instance. When the underlying server doesn't support running two concurrent instances, the sibling handshake fails.
In a normal chat session this failure is silently tolerated (falls back to "legacy" mode, server works fine). In Cowork/Code the identical failure is fatal and the server's tools become unavailable there.
First observed in logs on 2026-08-18. The server had worked in this exact configuration since June 2026 with no such error before that date, suggesting the sibling-probe mechanism is a recent addition.
I've also separately reported this to the Macuse maintainer, since the underlying fix (supporting concurrent instances, or responding gracefully when a second instance can't be spun up) is likely needed on their side. But I wanted to flag it here too since the error is fairly cryptic for a general user and it might be worth Claude Desktop failing more gracefully (e.g. falling back to the main-session instance) rather than silently dropping the server's tools in Cowork/Code.
Guida per i contributori
Nessuna guida per i contributori indicizzata per questo repository
Valutazione
Questa issue non è ancora stata valutata.