Claude Cowork: scheduled task gets HTTP 500 from a custom MCP (OAuth) connector, while Claude Desktop + Claude Code succeed on the same server

What happened

A Claude Cowork scheduled task ("weekly-briefing") received Tool call failed: 500
on every call to a custom remote MCP connector (OAuth, a Cloudflare Worker) — read
and admin tools alike, no usable payload.

At the same time, on the same MCP server:

• Claude Desktop (chat) → calls succeed, HTTP 200 with valid data
• Claude Code (CLI) → calls succeed, HTTP 200 with valid data

The exact tools that returned 500 in Cowork (e.g. admin_get_weekly_radar, sr_get_stats,
sr_get_articles) returned 200 from Desktop and Claude Code.

=> The failure is specific to the Cowork client/connector path, not the MCP server.

Evidence the MCP server was healthy throughout

• Same-window successful calls from Desktop and Claude Code (200 + valid payloads).
• Backend DB logs clean over the whole window (only 2xx, no connection errors, no restart).
• No worker deploy between the failure and recovery.

Environment

• Surface: Claude Cowork (claude.ai), scheduled task
• MCP server: custom remote OAuth connector (Cloudflare Worker)
• Window: 2026-06-06, ~07:30–08:30 UTC, self-resolved after ~1h
• Reproducibility: intermittent, no clear trigger

Impact

Scheduled task fails to import its data; output not generated. Discovered manually —
Cowork surfaced no error signal that distinguished it from a normal run.

Related

Same area as #61116 (Routines/agentic surface + MCP connectors). This manifests on
Cowork as a transient 500 rather than host_not_allowed.

Ask

Please check the Cowork → MCP connector / OAuth path for that window, and whether any
internal incident applies. Happy to provide the worker URL and timestamps privately.