Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report (please file separate reports for different bugs)
• I am using the latest version of Claude Code

What's Wrong?

Routines fail with "MCP tool call requires approval" despite connector panel guaranteeing no approval is needed

Summary

Claude Code Routines fail on every tool call to the Sentry connector with a "requires approval" error, even though the routine's Connectors panel explicitly displays the notice:

"Claude can use all tools from these connectors — including writes — without asking for permission during runs."

This is a direct contradiction between the routine UI's guarantee and runtime behavior.

Actual behavior

The first Sentry MCP tool call (find_organizations in my repro) fails immediately with:

Streamable HTTP error: Error POSTing to endpoint: MCP tool call requires approval

The agent then asks the user to approve permissions in chat — which is meaningless inside an unattended routine.

Environment

• Claude Code: desktop app (latest)
• Routine model: Opus 4.7 (1M context)
• Connector: Sentry, via https://mcp.sentry.dev/mcp (claude.ai connectors, not a plugin)
• Client OS: Windows 11 + WSL2 (routine itself runs in Anthropic cloud)

What I've ruled out

• Not a plugin/subagent issue — using the bare claude.ai connector, no Claude Code plugin installed.
• Not an OAuth scope issue — disconnected and re-authorized with every tool group ticked.
• Not a stale token / session state — fully quit Claude Code, fresh launch, recreated the routine.
• Not a Sentry server issue — identical connector works in a regular Claude.ai chat session (whoami returns the expected identity, find_organizations returns orgs).
• Not a permission allowlist issue — the routine's Connectors panel shows the explicit "no approval required during runs" notice.

Suspected cause

Looks like an MCP protocol elicitation/approval handshake that the routine runtime isn't auto-satisfying for this connector. Compare to openai/codex#11816, where calls hang/fail when an MCP server expects elicitation and the client can't respond. Either:

1. The routine runtime isn't auto-approving connector calls despite the UI promise, or
2. Sentry's MCP server is sending an approval/elicitation request that the routine runtime ignores and surfaces as this error.

Regardless of which side has the bug, the UI guarantee should match runtime behavior.

Screenshots

What Should Happen?

Per the notice in the Connectors tab, MCP tool calls to a configured connector should execute during a routine run without requiring approval. Routines run unattended and have no way to satisfy an approval prompt.

Error Messages/Logs

Steps to Reproduce

1. Connect Sentry at claude.ai/settings/connectors via OAuth, granting all tool group scopes.
2. Verify the connector works in a regular Claude chat session — whoami, find_organizations, etc. all succeed.
3. Create a routine with:
   • Any prompt that triggers a Sentry tool call (e.g., "Query Sentry for unresolved issues across all projects")
   • Sentry attached as the sole connector
   • Any trigger (schedule or Run now)
4. Trigger the routine.
5. Observe failure on the first Sentry MCP call.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

No response

Claude Code Version

Claude Code: desktop app (latest)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Warp

Additional Information

No response