Claude Code, the agent harness

Claude Code is Anthropic's official CLI agent. Claude, packaged with a harness that handles tool use, MCP, permissions, hooks, skills, and memory. It's the canonical starting point for autonomous work.

What it is, technically

A terminal application. You run claude in any directory; you get an interactive session with Claude (or automation-ready non-interactive mode). It has:

• A permission system for tool use
• Built-in tools (Read, Write, Edit, Bash, Glob, Grep, etc.)
• MCP client support (configure servers, call their tools)
• Hooks for event-driven automation
• Skills (named, invokable prompts)
• Persistent memory across sessions
• Sub-agent delegation
• Headless mode (for scripts, cron, CI)

Why it's the canonical harness

• It's maintained by the model maker. Tool-use changes in Claude ship with matching harness changes.
• The permission model is real. Not a checkbox, a structured allow/deny system that plays well with autonomy.
• MCP is first-class. Add an MCP server, it just works.
• Hooks and skills are composable. You can build sophisticated workflows without leaving the harness.
• Open format. Configs are JSON files you can check into a repo.

Install and first run

# Install
npm install -g @anthropic-ai/claude-code
# (or equivalent for your package manager)

# Run
claude

First run prompts you to authenticate. After that, you're in.

The surface area of this section

• Settings, the settings.json model
• Permissions, allow/deny, auto mode
• Hooks, event-driven automation
• Skills, named slash-commands
• Memory, persistent state

When to use Claude Code vs, a custom harness

Use Claude Code when:

• You want to move fast
• You're building personal productivity or internal tools
• You benefit from the ecosystem (MCPs, skills, plugins)

Build a custom harness when:

• You're embedding an agent inside a product you ship (users don't have terminals)
• You need tight control over every turn for evaluation purposes
• You're bridging multiple model providers and want a vendor-neutral layer