anthropics / anthropics/claude-code
[BUG] Custom remote MCP connector - all tools show "Auto-Allow disabled by your admin" in routine, despite org-wide Always Allow
- Dominant language
- Python
- Stars
- 145k
- Forks
- 23.1k
- PR merge metrics
- PR metrics pending
Description
### Preflight Checklist
- [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Abug) and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code
### What's Wrong?
Environment
Plan: Enterprise
Surface: Claude Code → Routines → routine detail (connector panel)
Connector: custom remote MCP server (org-managed, added via Organization settings > Connectors)
Observed
Every tool on a custom remote MCP connector shows "Auto-Allow disabled by your admin" in the routine's connector panel, and the routine cannot auto-approve them during unattended runs. This occurs despite all org admin settings being fully permissive:
- Org Settings > Connectors > [connector]: default restriction for unconfigured tools = Always allow; all tools "Uses default"
- Org Settings > Cowork > Permissions: "Allow 'Always allow' for connector tools" = on
- Org Settings > Claude Code > Managed settings: empty
- Routines: enabled
Expected
Per the routines docs, connector tools should run in unattended cloud sessions without approval prompts.
Notable / relation to closed issues:
- #61097 was closed as addressed but scoped its fix to Anthropic-hosted connectors (Gmail/Workspace/Notion), explicitly stating custom remote MCP connectors were unaffected. We are seeing the same class of failure on a custom remote MCP connector, which suggests either a regression or that the custom-connector path was never covered.
- #56757 (closed) reports the same "Always allow ignored" symptom.
Impact
Custom-connector routines are unusable for auto-publish/scheduled workflows; the documented core value of autonomous cloud execution with connectors does not hold for custom connectors.
Request
Confirm whether the custom-connector routine path is covered by the #61097 fix, and if not, treat this as the custom-connector variant.
### What Should Happen?
Setting auto-allow should be honored
### Error Messages/Logs
```shell
```
### Steps to Reproduce
Via claude:
Environment: Claude Code [X.Y.Z], macOS [version], Team plan.
Nine connectors are enabled for the workspace, including an
org-managed internal connector ("Connector A") and several
third-party ones. The workspace admin has disabled Auto-Allow on a
subset of tools across more than one connector.
1. Go to Routines > New routine. Scroll to the Connectors section
(shown by default — no tab switching needed). A banner reads:
"Claude can use all unrestricted tools from these connectors —
including writes — without asking for permission during runs."
and immediately below, in the same banner:
"Your admin restricts some tools in these connectors:"
followed by a list of tools each marked "Auto-Allow disabled by
your admin". The two halves contradict each other, and neither
states what happens at runtime when a restricted tool is needed.
Scope note: this banner appears only for cloud routines, not local
ones. The same connector and tool set works in a local session,
where a human can approve the prompt.
### Claude Model
None
### Is this a regression?
I don't know
### Last Working Version
_No response_
### Claude Code Version
N/A - Cowork
### Platform
Anthropic API
### Operating System
macOS
### Terminal/Shell
Terminal.app (macOS)
### Additional Information
_No response_
Contributor guide
No contributing guide indexed for this repository
Research direction
Start by reproducing the issue in Claude Code → Routines with an org-managed custom remote MCP connector and the stated Always Allow settings. Compare the behavior with closed issues #61097 and #56757. Done means custom-connector tools honor the configured auto-allow setting during unattended runs, without the contradictory restriction banner.
Written by the indexing model from the issue text.
Assessment
- Domain
- authorization, cloud
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Active
- Clarity
- Mostly clear
- Newbie friendliness
- 48/100