[BUG] Claude Code

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?

<img width="1300" height="1068" alt="Image" src="https://github.com/user-attachments/assets/39333724-ee5e-41c2-9f17-7b53cdcd87e8" />

Bug: False Usage Policy Violation When Reading Gmail via MCP in Agentic Loop
Environment
Claude Code Version: 2.1.5
OS: Windows 10/11
Model: claude-opus-4-5-20251101
MCP Servers: @gongrzhe/server-gmail-autoauth-mcp
Description
Claude Code throws a false "Usage Policy violation" error when autonomously reading emails through the Gmail MCP server during an agentic task execution. The error occurs mid-task without any user input, blocking the workflow entirely.

Error Message:

API Error: Claude Code is unable to respond to this request, which appears to violate our Usage Policy (https://www.anthropic.com/legal/aup). Try rephrasing the request or attempting a different approach. If you are seeing this refusal repeatedly, try running /model claude-sonnet-4-20250514 to switch models.

Steps to Reproduce
Configure Gmail MCP server (@gongrzhe/server-gmail-autoauth-mcp)
Enter Plan Mode and create a multi-step task involving email reading (e.g., "Build a workflow that reads my HARO emails and parses them")
Exit Plan Mode and let Claude begin autonomous execution
Claude successfully calls gmail: Read email once or twice
On subsequent automatic continuation (no user input), the Usage Policy error triggers
Expected Behavior
Claude should be able to read emails via an authorized MCP connection without triggering content policy violations, especially when:

The user has explicitly authorized the Gmail MCP integration
The user initiated the task requesting email access
The content being read is the user's own email
Actual Behavior
The content filtering system appears to flag the combination of:

Autonomous/agentic execution (no user message triggering the action)
Reading third-party content (email bodies)
Potentially sensitive topics in email content (journalist queries about health, legal, financial topics)
The filter cannot distinguish between "Claude is generating harmful content" vs "Claude is reading user-authorized external content."

Visual Evidence
The error occurs immediately after gmail: Read email tool calls, mid-execution, with no user input between the successful read and the error.

Impact
Blocks legitimate email automation workflows
Makes Gmail MCP integration unreliable for any multi-step task
Forces users to manually intervene at each step, defeating the purpose of agentic execution
Suggested Fix
Whitelist MCP tool responses from content policy scanning, or apply different thresholds
Distinguish between "Claude-generated content" and "externally-sourced content being processed"
Allow authorized MCP data sources to bypass content filters since the user explicitly granted access
Workarounds Attempted
Switching to Sonnet model (as suggested in error) - untested
Breaking tasks into manual single steps - works but defeats agentic purpose
Additional Context
This appears related to the broader challenge of AI assistants processing third-party content where content filters misattribute the source of potentially flagged material to the model rather than the external data source.

What Should Happen?

Claude should be able to skip emails or prompt the user before it shuts down the entire task

Error Messages/Logs

Steps to Reproduce

Steps to Reproduce
Configure Gmail MCP server (@gongrzhe/server-gmail-autoauth-mcp)
Enter Plan Mode and create a multi-step task involving email reading (e.g., "Build a workflow that reads my HARO emails and parses them")
Exit Plan Mode and let Claude begin autonomous execution
Claude successfully calls gmail: Read email once or twice
On subsequent automatic continuation (no user input), the Usage Policy error triggers

Claude Model

Opus

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.0.50 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Other

Additional Information

using desktop app