The agent-coding tool market stabilized around four shapes: terminal agents, IDE-integrated agents, VS Code extension agents, and lightweight pair-programmers. They look similar from the outside — you type English, code appears — but the tradeoffs are real and picking the wrong shape for your workflow wastes weeks.

This page compares the serious options as of April 2026. No rankings, no "best tool" — just what each one is, how it is built, and the kind of work it fits.

## The four tools worth knowing

| Tool | Shape | Primary loop | License |

| --- | --- | --- | --- |

| Claude Code | Terminal CLI + IDE extensions | Autonomous read-plan-edit-verify loop with permission model | Proprietary (Anthropic) |

| Codex CLI | Terminal, written in Rust | Autonomous agent with OS-level sandboxing | Open source (OpenAI) |

| Cursor Agent | Forked VS Code (full IDE) | Agent mode inside the editor with checkpoints | Proprietary |

| Cline | VS Code / JetBrains extension | Plan/Act modes with explicit approval on every change | Open source (Apache 2.0) |

A fifth tool, Aider, is worth mentioning for a specific niche: terminal-based, git-native, and deliberately small. It is the tool of choice if you want auto-commits per change and a small, inspectable loop over a large closed platform.

Claude Code is Anthropic's own agent, installed as a Node.js CLI (`npm i -g @anthropic-ai/claude-code`) and exposed through terminal, VS Code, and JetBrains extensions. It is the tool the rest of this wiki uses as its default reference.

**What is distinctive**

- **Permission-first architecture.** Every file edit, shell command, and network fetch is gated by a rule system that supports wildcards (`Bash(npm run *)`), path-scoped rules (`Edit(src/**)`), and modes (`default`, `acceptEdits`, `plan`, `auto`, `bypassPermissions`, `dontAsk`). See [Hooks and Policy-as-Code](/s/agents-development/wiki/hooks-policy-as-code) for how to enforce what you cannot express declaratively.

- **First-party MCP support.** MCP is not a plugin — it is the default way to connect external tools. Scopes are `local`, `project` (checked into `.mcp.json`), and `user`. Covered fully in [MCP](/s/agents-development/wiki/mcp).

- **Subagents and headless mode.** Claude Code ships with built-in subagents (Explore, Plan, general-purpose) and lets you define custom ones as markdown files in `.claude/agents/`. For CI/CD, `claude -p "task"` runs non-interactively and supports `--output-format json`. Both covered in [Multi-Agent Patterns](/s/agents-development/wiki/multi-agent-patterns).

- **CLAUDE.md is the rules system.** Instructions live in plain-text markdown files with a clear precedence. See [CLAUDE.md and Agent Rules](/s/agents-development/wiki/claude-md-agent-rules).

- **Enterprise-ready settings.** Managed policy files on macOS/Linux/Windows override user settings and cannot be bypassed. See [Enterprise Rollout](/s/agents-development/wiki/enterprise-rollout).

**Best for**: teams who want a terminal-first agent with auditable guardrails, and who can invest in writing CLAUDE.md and hooks. Strongest when you treat the agent as a governed colleague, not a prompt box.

**Weakest at**: tight in-editor refactoring flows where you want inline diff + accept. The VS Code and JetBrains extensions help but are not the primary surface.

Sources: [docs — overview](https://code.claude.com/docs/en/overview), [permissions](https://code.claude.com/docs/en/permissions), [headless](https://code.claude.com/docs/en/headless), [memory](https://code.claude.com/docs/en/memory).