A platform using MCP implements a tool server with 47 distinct tools across multiple domains. Claude agents using this server sometimes call the wrong tool due to similar tool names (e.g., send_email vs. queue_email) or vague descriptions. This causes incorrect side effects in production. What MCP schema design practices reduce tool confusion?

MCP tool descriptions are Claude's primary signal for tool selection. Well-crafted descriptions prevent misuse by making the tool's scope, behavior, and side effects unambiguous. Differentiating send_email ('immediately sends an email via SMTP, cannot be recalled after execution') from queue_email ('adds to the outbound queue for batch delivery in the next scheduled run, can be removed before the run') in the description eliminates the ambiguity. Naming conventions, such as prefixing irreversible tools, provide an additional heuristic signal. The approach that reduces the total number of tools to under 10 by combining related tools into multi-function tools with a 'mode' parameter trades tool confusion for mode confusion — email_tool(mode='send') vs. email_tool(mode='queue') has the same disambiguation problem with a different UI. The approach that implements a tool disambiguation layer where Claude calls a which_tool meta-tool with its intent adds a round-trip and doesn't solve the underlying description quality problem — Claude would still need to understand tool differences to answer the meta-tool query. The approach that restricts Claude to 5 tools per request context and dynamically injects only the relevant subset based on the current task is a valid scaling strategy for 47+ tools but doesn't address the description quality issue.