[BUG] MCP Tool Permissions Not Persisting Between Chats in claude.ai Projects (Team Plan)

Preflight Checklist

• [x] I have searched existing issues 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?

MCP tool permissions configured as "Always allow" in Settings → Connectors do not persist between different chats within the same Project.

Scope:

• ✅ Normal chats (outside Projects): Permissions persist correctly
• ❌ Chats inside Projects: Permissions reset with every new chat

Every new chat within a Project requires re-approving MCP tool permissions, even though:

• The connector is configured as "Always allow" in Settings
• The same user already approved the tools multiple times
• Normal chats (outside Projects) work without prompts

Note: This affects claude.ai web interface (not Claude Code CLI), specifically within Projects on Team Plan.

What Should Happen?

When a user configures "Always allow" permissions for an MCP connector in Settings:

1. The permission should persist across all chats (both in and out of Projects)
2. Once approved, subsequent chats in the same Project should not prompt again
3. The behavior should be consistent between Project chats and normal chats

Error Messages/Logs

N/A - No error messages, just incorrect permission behavior

Steps to Reproduce

Setup:

1. Log into claude.ai with a Team Plan account
2. Navigate to Settings → Connectors
3. Configure a custom MCP server (e.g., "Example Internal MCP")
4. In the "Tool permissions" dropdown, select "Always allow"

Test outside Projects (works correctly):

1. Open a new chat (not in a project)
2. Use any MCP tool → Works without prompts ✅
3. Open another new chat → Still works without prompts ✅

Test inside Projects (bug occurs):

1. Navigate to a Project
2. Open a new chat within the Project
3. Use any MCP tool → Permission prompt appears ❌
4. Select "Allow always" (Ctrl + Enter)
5. Tool works in current chat
6. Open another new chat in the same Project
7. Use the same MCP tool → Permission prompt appears AGAIN ❌

Claude Model

Sonnet (default)

Is this a regression?

I don't know

Last Working Version

N/A

Claude Code Version

N/A - This bug affects claude.ai web interface, not Claude Code CLI

Platform

Other

Operating System

Other

Terminal/Shell

Terminal.app (macOS)

Additional Information

Impact on Team Plan Users

This significantly impacts Team Plan users who rely on MCP tools for collaborative work in Projects. Team members must re-approve permissions for every single new conversation.

Related Issues

This appears to be the web/Project-scoped variant of existing permission bugs:

• #7472 - "Always allow from this project" not persisted (CLI)
• #5140 - Permissions from user settings not applied at project level (CLI)
• #3107 - MCP wildcard permissions not honored (CLI)
• #14730 - Wildcard Permissions Matching Fails (Dec 2025, CLI)
• #5749 - MCP Permission System Broken in Web Interface

Technical Notes

• Account Type: Paid Team Plan with organization-level MCP connectors
• MCP Server Type: Custom remote MCP server
• Configuration: Added via Admin Settings → Connectors → Add custom connector
• Tested By: Multiple team members experiencing same issue
• Reproducibility: 100% reproducible on every new Project chat