- Tool dispatch: Claude receives a list of available tools in its system prompt and names each one in a tool_use block along with structured inputs.
- Schema validation: The harness validates each tool call against its input schema before execution.
- Permissions: Tools that write or execute are gated by permission rules (allow/ask/deny) defined in
settings.json or via /permissions.
- Tool classes: Tools are categorized by side-effect type: read-only (Glob, Grep, Read), writes (Write, Edit), executes (Bash, PowerShell), spawns (Agent, TaskCreate), orchestration (SendMessage, TeamCreate), or meta (Config, ToolSearch).
- Gating mechanisms: Some tools are always on; others are entrypoint-gated (SDK-only, local-agent-only, remote-only), feature-flag-gated (env vars like
CLAUDE_CODE_ENABLE_TASKS), or subscription-gated (Pro/Max).
- MCP orchestration: The four MCP tools (ListMcpResources, ReadMcpResource, plus indirect orchestration via Tool Search) are primitives for discovering and using dynamically-registered tools from MCP servers; individual MCP tools themselves are not built-in.
- Deferred loading: ToolSearch (when
ENABLE_TOOL_SEARCH=1) loads tool schemas lazily to reduce initial system prompt size; MCP tools are also deferred until needed.
← Back to Tools/README.md
