MCP tools are typed, described functions an AI model can call — the unit of capability that decides whether an agent is useful or dangerous.
1 min readupdated 2026-08-01
/ quick answer
Each MCP tool has a name, a description, a JSON-schema input and annotations (read-only, destructive, idempotent). The model chooses tools from those descriptions alone, so tool design is prompt design: narrow scope, unambiguous names, small inputs and compact outputs. Write tools should carry approval requirements.
MCP tools are typed, described functions an AI model can call — the unit of capability that decides whether an agent is useful or dangerous. Each MCP tool has a name, a description, a JSON-schema input and annotations (read-only, destructive, idempotent). The model chooses tools from those descriptions alone, so tool design is prompt design: narrow scope, unambiguous names, small inputs and compact outputs. Write tools should carry approval requirements. In practice: `search_orders(customer_email, status)` returning 10 compact rows beats a generic `run_sql(query)` tool that hands the model the whole database. This dictionary node is part of the Onexial knowledge graph and links to related concepts, workflows and tools below.
Definition
Each MCP tool has a name, a description, a JSON-schema input and annotations (read-only, destructive, idempotent). The model chooses tools from those descriptions alone, so tool design is prompt design: narrow scope, unambiguous names, small inputs and compact outputs. Write tools should carry approval requirements.
Example
`search_orders(customer_email, status)` returning 10 compact rows beats a generic `run_sql(query)` tool that hands the model the whole database.
Each MCP tool has a name, a description, a JSON-schema input and annotations (read-only, destructive, idempotent). The model chooses tools from those descriptions alone, so tool design is prompt design: narrow scope, unambiguous names, small inputs and compact outputs. Write tools should carry approval requirements.
What is an example of MCP Tools?
`search_orders(customer_email, status)` returning 10 compact rows beats a generic `run_sql(query)` tool that hands the model the whole database.
Why does MCP Tools matter for AI and automation?
MCP tools are typed, described functions an AI model can call — the unit of capability that decides whether an agent is useful or dangerous. It connects to the workflows, prompts and tool stacks linked on this page, so you can move from definition to execution without leaving Onexial.