Claude Code¶

Tool-specific reference for Claude Code's agentic features.

These features run on Anthropic's current model line. As of June 2026 the Claude models reference lists Claude Fable 5 at general availability and Claude Mythos 5 in preview, both fronting adaptive thinking and a 1M-token context window — positioned above the prior Opus 4.8 generation (Claude models overview).

Pages¶

Sub-Agents — Ephemeral, isolated agents for focused task execution
Agent Teams — Coordinated multi-agent sessions with shared task lists
Agent View — Dispatch-attach-monitor surface for parallel background sessions, with blocked-on-input as a first-class state
Auto Mode — Classifier-based permission gating for unattended execution
Hard-Deny Classifier Rule — Unconditional block layer inside auto mode's classifier — user intent and allow exceptions do not apply
Parameter-Level Permission Rules — Tool(param:value) syntax matches deny/ask rules against a tool's input-parameter values, e.g. Agent(model:opus) to block Opus subagents
Bare Mode — Deterministic headless CI mode — skip all local config discovery with --bare
Extension Points — Decision framework for choosing between CLAUDE.md, rules, skills, hooks, subagents, MCP, and plugins
Hooks & Lifecycle — Deterministic automation at 12+ lifecycle events
Agent SDK — The Claude Code runtime as a library for custom applications
/batch & Worktrees — Parallel execution with worktree isolation
Feature Flags & Environment Variables — Curated reference for the most impactful configuration knobs
Session Scheduling — /loop and cron tools for recurring prompts within a session
Cloud-Scheduled Routines vs Local Session Scheduling — When to move schedule out of the local box; cloud Routines trade working-tree fidelity for uptime continuity
Skill Eval Loop — Test, benchmark, A/B-compare, and optimize agent skills with the skill-creator eval framework
Reloading Skills Mid-Session — Re-scan skill directories with /reload-skills or a SessionStart hook, picking up edits without losing context
Monitor Tool — Stream stdout from background scripts to Claude line-by-line, eliminating polling loops
Plugin Background Monitors — Declarative supervision auto-armed at session start via the monitors manifest key
Plugin-Activated Main-Agent Override and Bin/ PATH Injection — Plugin settings.json swaps the main thread agent; bin/ injects executables onto the Bash tool's PATH for the plugin's enabled lifetime
Channels Permission Relay — Forward tool-use approval prompts to your phone for unattended agent runs
Managed Settings Drop-In Directory — Deploy independent policy fragments per team without merge conflicts
PowerShell Tool — Native Windows shell for Claude Code via PowerShell instead of Git Bash
Skill disallowed-tools Frontmatter — Skill-layer tool denial: remove tools from the model while a skill is active, the deny-side complement to allowed-tools
Local Plugin Scaffolding (claude plugin init) — Auto-load plugins from .claude/skills/<name>/.claude-plugin/plugin.json and scaffold the layout in one command — when the manifest earns its keep over a loose skill, and when it does not