issue with MCP integration in Claude.ai

### What happened?

[BUG] Custom StreamableHTTP MCP server intermittently fails to load in Claude Desktop after updates

## Preflight Checklist
- [x] I have searched existing issues and this hasn't been fully resolved
- [x] This is a single bug report
- [x] I am using the latest version of Claude Desktop

## Summary

A custom self-hosted MCP server (StreamableHTTP, Bearer Token auth) intermittently fails to load its tools in Claude Desktop after Anthropic updates. The failure is **not** caused by network issues, Cloudflare tunnel problems, or server-side configuration — confirmed by the fact that **Opencode and Codex connect to the same MCP server on the same machine without any issues** throughout the same period.

## Environment

| Field | Detail |
|---|---|
| Client | Claude Desktop (latest) |
| OS | Windows / macOS |
| MCP Server | Self-hosted Gitea instance |
| Transport | StreamableHTTP |
| Auth | Bearer Token |
| MCP Server URL | https://gitea-mcp.rtsb.ai/mcp |

## What's Wrong?

Every few days — consistently correlated with Claude Desktop updates — the custom MCP connector stops loading its tools. Claude Desktop shows the connector as configured but the tools are not available in session.

The only workaround discovered so far: switching to a different model (e.g. Fable) sometimes restores tool availability. This is not a reliable or acceptable fix.

## Key Evidence

- **Same machine, same network, same MCP server:** Opencode and Codex both connect and use the MCP server without any issues during the exact same period that Claude Desktop fails.
- **Not a network/tunnel issue:** No VPN changes, no Cloudflare tunnel restarts, no firewall changes — nothing on the infrastructure side changes between working and broken states.
- **Not user error:** The connector config in Claude Desktop is unchanged and valid.
- **Reproducible pattern:** Issue appears reliably after Claude Desktop updates, resolves temporarily, then recurs after the next update.
- **Historical workaround:** In past sessions, calling `tool_search` with query `"gitea"` at session start would surface the deferred tools — but this internal mechanism is no longer consistently accessible from within Claude Desktop sessions.

## Expected Behavior

A configured and reachable MCP server should load its tools reliably after every Claude Desktop update, without requiring workarounds or model switching.

## Steps to Reproduce

1. Configure a custom Streamable HTTP MCP server in Claude Desktop with Bearer Token auth
2. Update Claude Desktop to a new version
3. Open a new session — MCP tools fail to load despite server being reachable
4. Confirm server is reachable from the same machine via other clients (Opencode, Codex, curl)

## Impact

This is a production blocker for users relying on custom MCP integrations as part of their core workflow. Recurring regressions after every update make the tool unreliable for professional use.

### What did you expect to happen?

A configured and reachable custom MCP server (StreamableHTTP + Bearer Token) should load its tools reliably in every Claude Desktop session, including after app updates — without requiring workarounds such as model switching or session restarts.

### Steps to reproduce

1. Configure a custom StreamableHTTP MCP server in Claude Desktop
(Bearer Token auth, e.g. https://gitea-mcp.rtsb.ai/mcp)
2. Anthropic releases a Claude Desktop update — app auto-updates
3. Open a new session
4. MCP tools fail to load despite server being reachable
5. Verify: same server accessible from Opencode and Codex
on the same machine without issues

### Area

MCP Connector (adding/managing servers)

### MCP Server (if applicable)

_No response_

### Error messages or logs

```shell

```

### Additional context

_No response_

No repository file, test, or entry point is identified in the report. Start by reproducing the failure with the listed StreamableHTTP and Bearer Token setup after a Claude Desktop update, compare connectivity with curl, Opencode, and Codex, and collect Claude Desktop logs; done means the regression is isolated and a reliable fix is identified.

Vom Indexierungsmodell aus dem Issue-Text verfasst.