Tools Landscape

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 but the tradeoffs are real.

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

The four tools worth knowing

Tool	Shape	Primary loop	License
Claude Code	Terminal CLI + IDE extensions	Autonomous loop with strong 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 with checkpoints	Proprietary
Cline	VS Code / JetBrains extension	Plan/Act with approve-every-change	Open source (Apache 2.0)

Aider is worth mentioning for a specific niche: terminal-based, git-native, deliberately small.

Claude Code

Anthropic's agent, installed via npm i -g @anthropic-ai/claude-code and exposed through terminal, VS Code, and JetBrains.

• Permission-first architecture. Every action is gated by a rule system supporting wildcards (Bash(npm run *)), path-scoped rules (Edit(src/**)), and modes: default, acceptEdits, plan, auto, bypassPermissions, dontAsk. See Hooks and Policy-as-Code.
• First-party MCP. Scopes: local, project (.mcp.json), user. See MCP.
• Subagents and headless mode. Built-in subagents (Explore, Plan, general-purpose) plus custom ones in .claude/agents/. For CI, claude -p "task" --output-format json runs non-interactively. See Multi-Agent Patterns.
• CLAUDE.md. Plain-text markdown rules with a clear precedence. See CLAUDE.md and Agent Rules.
• Enterprise settings. Managed policy files override user settings and cannot be bypassed. See Enterprise Rollout.

Best for: terminal-first teams who want auditable guardrails and can invest in CLAUDE.md and hooks.

Weakest at: inline-diff IDE refactoring flows.

Sources: docs, permissions, headless.

OpenAI Codex CLI

OpenAI's open-source agent, written in Rust, installed with npm i -g @openai/codex.

• Rust + OS sandbox. Apple Seatbelt on macOS, Landlock + seccomp on Linux. The sandbox is the enforcement layer.
• Approval modes. Configurable policies for when edits need sign-off.
• Models. GPT-5.4, GPT-5.3-Codex, others with reasoning levels.
• Subagents and MCP. Task parallelization and external tool integration.
• Image inputs. Screenshots and design specs as context.

Best for: teams in the OpenAI ecosystem, people who trust OS sandboxing over prompt-based permissions.

Weakest at: Windows is experimental (WSL recommended).

Sources: Codex CLI docs, GitHub.

Cursor

A fork of VS Code with the agent built into the IDE.

• Full IDE, not an extension. Inline diffs, checkpoints, agent state tightly integrated with the editor.
• Checkpoints. Codebase snapshots per session — safety net for long autonomous runs.
• Queued messages. Stack instructions while the agent works.
• Tool-agnostic models. Anthropic, OpenAI, and others, each with tuned tool prompts.
• MCP support plus a rules system (.cursorrules) analogous to CLAUDE.md.

Best for: developers who want the agent deeply integrated with an editor they already live in.

Weakest at: headless CI and reproducibility — a lot of the power lives in the UI.

Sources: Cursor Agent docs.

Cline

Open-source VS Code, Cursor, and JetBrains extension (Apache 2.0) built around explicit user approval for every action.

• Approve-every-change. Every edit and command surfaces a confirmation. Slow but trivially auditable.
• Plan/Act modes. Plan explores, Act executes. The separation is explicit.
• Bring-your-own-model. Anthropic, OpenAI, Google, OpenRouter, DeepSeek, Ollama.
• Client-side only. Context flows directly to the model provider.
• MCP support.

Best for: regulated environments, learners who want to see every step, teams on strict data paths.

Weakest at: speed.

Sources: Cline docs.

Aider (honorable mention)

A terminal-based pair programmer. Maps your whole codebase, makes one change, auto-commits it. You review with normal git tools.

• Auto-commits per change.
• Codebase map for larger repos.
• Claude, GPT-5.4, DeepSeek, local models.
• Voice-to-code, image context, linting + tests.

Best for: solo developers who want a small, inspectable tool with git as the audit log.

Source: aider.chat.

Picking one

Do not pick by "which is best". Pick by workflow:

• Terminal-first, CI, enterprise guardrails: Claude Code
• Open source, sandbox over prompt rules: Codex CLI
• IDE-native, inline diffs: Cursor
• Regulated, approval per change: Cline
• Git-native, small surface: Aider

A common combination: Cursor for editor work plus Claude Code for terminal and CI. They read the same CLAUDE.md and share MCP servers.

Benchmarks — skip them

No honest public benchmark will tell you which tool fits your codebase. Run a real task from your own backlog in each tool for an hour. That is the benchmark.

---

Next: CLAUDE.md and Agent Rules

Sources

• Claude Code docs
• OpenAI Codex CLI
• Cursor Agent
• Cline
• Aider
• MCP