Claude.ai hosted MCP connectors drop sessions frequently with no auto-reconnect

Problem

The Claude.ai hosted MCP connectors (PubMed, Clinical Trials, Consensus, Gmail, Google Calendar, etc.) frequently drop sessions mid-conversation with no automatic reconnect. The error is:

Streamable HTTP error: Error POSTing to endpoint: 
{"type":"error","error":{"type":"not_found_error","message":"MCP session has been terminated or no longer exists on the server."},"request_id":"..."}

This happens unpredictably — sometimes on the first call, sometimes after a few successful calls in the same session. There's no retry logic or session recovery. Once the session drops, subsequent calls to the same MCP server also fail.

Impact

• Tools that depend on these connectors become unreliable for any multi-step workflow
• Skills/prompts that chain MCP calls (e.g., PubMed search → get article metadata → get full text) frequently fail partway through
• Users can't build reliable automation on top of hosted MCP tools when session stability is unpredictable
• The workaround is building self-hosted MCP servers wrapping the same public APIs (PubMed E-utilities, ClinicalTrials.gov v2), which defeats the purpose of offering hosted connectors

Expected Behavior

• Sessions should auto-reconnect transparently on transient failures
• If a session is truly terminated, the client should establish a new session and retry the failed call
• At minimum, a clear retry mechanism or session refresh tool should be available

Environment

• Claude Code CLI (claude-opus-4-6 with 1M context)
• macOS Darwin 25.3.0
• Observed across multiple MCP servers: PubMed, Clinical Trials, Consensus
• Reproducible across sessions and days — not a one-off