known_issues

Known Issues

Unintended Command Execution by the AI Assistant

In some cases, the MCP server can execute unintended or incorrect commands. This usually happens when the assistant is working with a lot of old or outdated conversation history (called “context”), which can confuse it or cause it to misunderstand what you are asking.

This issue is related to a known behavior of large language models (LLMs), where they can "hallucinate" — that is, generate outputs that seem correct but are actually wrong or not intended.

How to Reduce the Risk

To improve the reliability of the assistant and avoid unwanted system actions, we recommend:

• Avoid long conversations: Don’t keep a single chat going for too long. Start a fresh session or new chat if you're making a separate request.
• Keep instructions clear and simple: Make sure your request is easy to understand. Break complex tasks into smaller steps.
• Double-check commands: Before running any command suggested by the assistant, review it carefully to ensure it does what you expect.
• Use recent context only: If your client allows it, limit how much past conversation is included. Old information can confuse the assistant.
• Be specific with instructions: The more precise you are, the lower the chance of misunderstandings.
• Restrict write access: For sensitive or irreversible operations, we strongly recommend starting the server with the --no-allow-write option. This prevents the assistant from directly executing write or destructive commands, adding an extra layer of safety.