A team's MCP server had 12 tools. Names like do_thing, process, helper. The AI assistant called the wrong tool 30% of the time.
The fix: rename. Tools became extract_invoice_line_items, summarise_meeting_notes, find_customer_by_email. Wrong-tool rate dropped to under 5%.
Tool names are the first thing the AI assistant sees. Naming matters.
The naming rules
Each tool name should:
- Start with a verb (extract, summarise, find, send).
- Specify the object (invoice line items, meeting notes, customer).
- Be unique enough that the AI doesn't confuse it with siblings.
- Avoid abbreviations the AI may not parse correctly.
Reviewer ritual
PR review for new tools:
- Name passes the rules.
- Name doesn't conflict with sibling tools.
- Name reads naturally to someone who hasn't seen the tool before.
A real naming guide
A team's MCP server naming standards:
verb_object_qualifierpattern.- Examples:
get_customer_by_id,list_invoices_for_customer,send_notification_to_user. - Avoid:
process_*,handle_*,*_thing.
Trade-offs
Long names vs. short names:
- Long: more discoverable, more to type in tests.
- Short: faster but ambiguous.
Lean long. The AI assistant cares about clarity, not character count.
Limits
Some operations don't fit cleanly:
- "Search across multiple tables" — what's the object?
- "Multi-step workflows" — better as multiple tools.
When in doubt, split.
What we won't ship
Tool names that don't follow the rules.
Tools with names that conflict.
Tools named for engineering convenience rather than AI clarity.
Renames without coordinated migration plans.
Close
MCP tool naming is the simplest discipline that produces measurable wins. Verb + object + qualifier. Unique, clear, descriptive. Skip the discipline and the AI guesses.
Related reading
- Tool design like APIs — surrounding discipline.
- MCP tool schemas — companion topic.
- Your first MCP server — surrounding context.
We build AI-enabled software and help businesses put AI to work. If you're tightening tool naming, we'd love to hear about it. Get in touch.