A Tool in 📝Claude Code is a typed capability the agent can invoke during a session — built-in (Read, Write, Edit, Bash, etc.), MCP-provided, or skill-provided — each governed by the permissions system before execution.
Built-in tools ship with the Claude Code binary and cover file operations (Read, Edit, Write), shell execution (Bash), web access (WebFetch, WebSearch), task management (TaskCreate, TaskUpdate, TaskList), background work (Monitor), planning (EnterPlanMode, ExitPlanMode), and agent orchestration (the Agent tool for spawning subagents). Each tool exposes a JSONSchema for its inputs, and the harness validates calls before dispatch.
📝MCP Server tools register dynamically when an MCP server connects — they appear in the model's tool catalog prefixed with mcp__<server>__<tool> (e.g., mcp__mythos__read_memo). MCP tools are interchangeable with built-ins at the invocation layer; the only structural difference is the prefix and the connection lifecycle. Skills can also expose tools, though more commonly they shape the model's behavior via system-prompt injection rather than registering new callable surfaces.
Tool availability per session is governed by 📝Permissions (Claude Code) declared in 📝settings.json (Claude Code) — each call passes through allow / ask / deny resolution before execution. Some tools are also gated by the session's permission mode (accept-edits, plan, default), which can constrain mutating operations regardless of explicit permission rules.