14days-build-claude-code-cli

Skill by ara.so — Claude Code Skills collection.

A 14-day tutorial project teaching you to build a production-style AI code agent CLI (agent-code) in Python. Learn the harness architecture: tool calling, file operations, bash execution, permissions, session memory, hooks, skills, subagents, worktree isolation, and MCP integration.

What This Project Does

This is an educational implementation of a Claude Code-style agent harness. You'll build agent-code, a CLI that:

• Accepts one-shot prompts or enters an interactive REPL
• Calls real LLMs (default: DeepSeek via Anthropic-compatible endpoint)
• Uses tool calling (tool_use / tool_result) with file operations, web search, and bash
• Implements safety: read-before-edit, diff preview, permission system
• Manages sessions, project memory, and context compression
• Supports slash commands, hooks, skills, subagents, and MCP tools

Core philosophy: The model handles reasoning; the harness handles context, tools, permissions, execution, state, and feedback loops.