Schedule skill / RemoteTrigger API doesn't see MCP connectors connected on Team workspace

Seeing a related issue this morning on https://claude.ai/code/routines/new

<img width="438" height="240" alt="Image" src="https://github.com/user-attachments/assets/62f95c2c-9523-4636-a0e9-c381914a1695" />

Meanwhile in settings:

<img width="1279" height="519" alt="Image" src="https://github.com/user-attachments/assets/b91b7848-dd0f-4f4a-8429-acd3fca59def" />

Workaround that works today (Team plan, Claude Code Desktop + claude.ai web, same account)

Confirmed this bug on Claude Code 2.1.150, Windows 11, Team workspace. Slack connector is visible and functional at claude.ai/customize/connectors but the /schedule skill returns No connected MCP connectors found and the routines API returns an empty connector list.

Working bypass: Hardcode the connector UUID and URL directly into mcp_connections when calling RemoteTrigger — do not rely on the skill's connector-discovery output at all.

"mcp_connections": [
{
"connector_uuid": "5df403b0-baf3-4648-bebf-b2ee2fea3fbf",
"name": "Slack",
"url": "https://mcp.slack.com/mcp"
}
]
How to find your UUID: Even though the skill's connector list is empty, the UUID is discoverable — call RemoteTrigger { action: "list" } on any existing routine that was successfully created with a connector attached, and read the mcp_connections array from the response. The UUID and URL are both present there.

Once hardcoded into mcp_connections on routine creation, the connector attaches and works correctly — Slack messages send successfully from the remote CCR agent.

Confirming the hypothesis: The UUID exposed in local tool names (e.g. mcp__5df403b0-... in the Desktop session) does not appear in the routines API connector list, consistent with the individual-user vs. workspace scope boundary described above.

This workaround defeats the purpose of automated connector discovery and requires users to already know the UUID, but it does unblock routine creation until the scope issue is resolved.

Hitting this today on Claude Code Desktop. One nuance worth adding:

The /schedule skill reports "No connected MCP connectors found" — same as the original report. But when I created a routine via RemoteTrigger create without specifying mcp_connections, the API auto-attached all 8 of my connected connectors (Linear, Slack, Gmail, Google Drive, Calendar, Read.ai, Enterpret, Zapier).

So it looks like the API layer can see my connectors fine — it's the connector inventory the /schedule skill displays to the user that's returning empty. Two layers, two different bugs (or one bug with an inconsistent surface).

User impact is unchanged from the original report: the skill output strongly implies "your connectors aren't connected, go fix that," which sent me on a 20-minute troubleshooting loop (reconnecting connectors, restarting sessions) that turned out to be unnecessary.

Environment: Claude Code Desktop, individual account (not Team workspace), so this isn't specific to Team scope as the original report hypothesizes.

Still happening as of 2026-06-16 with a new failure mode — the documented subagent workaround from #43397 no longer helps.

Minimal repro:

• Create a scheduled routine via RemoteTrigger, with mcp_connections populated for the connectors you want.
• Set session_context.allowed_tools to explicitly include the mcp__<Server> patterns for those connectors.
• Prompt: spawn a subagent that runs ToolSearch with query "mcp" (max_results=100) and reports back the list of MCP tools it sees. No actual MCP calls are made.
• Result: 0 MCP tools visible to the subagent.

Reproduces in two environments I have access to (one custom, one Default) — both FAIL. Same OAuth credentials work fine in a local Claude Code session for the same user on the same day, so this is not a connector-side auth issue.

Real impact: four scheduled routines on my account have been silently dead since 2026-06-08 (~14:00 UTC). The subagents in those runs report "no MCP tools available" and exit successfully. The runs UI shows them as completed; no side-effects on any of the configured systems are observable. Internal fail-loud / RED-ALERT mechanisms can't fire because they rely on the same connectors that are missing.

The #43397 workaround (spawn a subagent so it gets a fresh tool context) no longer helps — the subagent also sees zero MCP tools.

Happy to share account / routine IDs privately if useful for triage.

Fresh repro on the current build filed as #73845 — CC 2.1.199, remote_trigger entrypoint. ListConnectors from inside the routine shows every connector connected: true but enabledInChat: false, so no mcp__* tools are ever injected. The subagent-dispatch workaround (#43397) and toggling the connector on in the routine's web-UI Connectors tab both fail now — top-level session and subagent alike see zero connector tools. Looks like a regression of the #35899 fix (v2.1.101) at the routine entrypoint.