The Tool Description Problem
The number one reason your AI assistant ignores a connected MCP server is that the tool descriptions don't match how you phrase your requests. You ask "check if there are any error logs from today" and the model doesn't connect that to your logging MCP server because the tool is described as "query structured log data using LogQL syntax."
The model matches your natural language request against the tool descriptions it received when the servers connected. If there's a vocabulary mismatch, the model won't make the connection. It's not being stupid. It's being literal about matching your intent to available capabilities.
How to Fix Description Mismatches
If you control the MCP server, improve the tool descriptions. Add natural language examples of when the tool should be used. Instead of just "Execute a database query," write "Execute a SQL query against the connected database. Use this when the user asks about data, records, counts, trends, or any question that could be answered by querying a relational database."
If you don't control the server, you can work around the issue by being more explicit in your prompts. Instead of "check the logs," try "use the logging tool to search for error entries from today." Naming the tool or capability directly gives the model the hint it needs.
Context Competition
When you have many tools connected, they compete for the model's attention. The model has a limited "working set" of tools it's actively considering for each request. Tools with clearer, more specific descriptions tend to win this competition. Generic tools with vague descriptions get overlooked.
This is why tool selection matters. If two tools could plausibly handle a request, the model picks one based on description clarity and relevance. The loser doesn't get used, even though it might have been the better choice.
The Familiarity Bias
Models develop a kind of familiarity bias during a conversation. Once they've successfully used a tool, they're more likely to use it again for similar requests. If the model used your database server to answer one question, it'll tend to reach for the database server again even when a different tool would be more appropriate.
You can break this pattern by explicitly mentioning the tool you want used. "Using the file search tool (not the database), find all configuration files in the project" overrides the model's default tendency and forces it to consider the tool you specified.
Too Many Similar Tools
If you have two database MCP servers connected, or a file search server alongside a code search server, the model might struggle to differentiate them. It picks one consistently and ignores the other. The fix is to either disconnect the redundant server or ensure their descriptions clearly differentiate their purposes.
Related Reading
- How AI Assistants Choose Which Tool to Use
- How to Debug an AI Agent That Keeps Making Mistakes
- Setting Up Claude with MCP Servers for Daily Work
Browse MCP servers on Skillful.sh. Search 137,000+ AI tools.