As of 2026-05-21

As of 2026-05-21

There are now enough AI coding tools that picking one feels like its own evaluation problem. The good news is that the landscape has stabilized into three shapes. Once you pick the shape that fits your habits, the specific tool inside that shape is a much smaller decision.

The three shapes

Terminal-native agents. Run in your shell, edit files in your repo, run commands, iterate. The model lives behind a CLI, not behind a UI panel. Best for engineers who already live in a terminal and tmux/screen all day. Representative tools: Claude Code, Aider, OpenAI Codex CLI.

IDE-first environments. Either a fork of VS Code (or another editor) with AI built in, or an editor designed from scratch around AI. The agent is wired tightly into your existing edit/file/diff experience. Best for engineers who want a single tool that handles everything. Representative tools: Cursor, Windsurf (Codeium's editor), Zed AI.

Editor extensions and copilots. Add AI assistance to your existing editor without replacing it. Lighter weight than a full IDE switch, less agentic than a terminal agent. Best for teams that want incremental AI integration into an established workflow. Representative tools: GitHub Copilot (including Copilot Workspace), Cline (VS Code agent), Continue, JetBrains AI Assistant.

Once you know which shape you want, the choice within it is mostly about model access, pricing, and how aggressively agentic the tool tries to be.

The current major tools

Claude Code — Anthropic's terminal CLI. Repo-aware, agentic, optimized for the Claude models. Best at multi-file edits, fast edit-test loops, and "stay in the terminal" workflows. Plugs into VS Code and JetBrains via extensions but the heart of the tool is the CLI.

Cursor — A VS Code fork built around AI. Inline edit/completion, multi-file chat, an agent mode that can take a goal and execute it across files. Strong at the IDE experience; well-suited if you would have used VS Code anyway. Subscription model with a free tier and paid pro tier.

Aider — Open-source CLI, model-agnostic. Pair-programming-shaped, very transparent about every change it makes via git. Works with Claude, OpenAI, DeepSeek, Gemini, and local models. The most "engineer's tool" of the bunch — config-driven, repo-aware, no SaaS lock-in. See how-to-use-aider-effectively for a deep dive.

OpenAI Codex / Codex CLI — The naming has shifted several times. As of 2026 the relevant thing is OpenAI's terminal-based coding agent (Codex CLI) plus an API surface for its coding-tuned models. Best when you want OpenAI models specifically and a terminal-native experience.

GitHub Copilot and Copilot Workspace — The grandparent of the category. Copilot for inline completion is in the editor; Copilot Workspace adds an agentic "plan and edit" surface that lives in GitHub itself. Strong fit if you are already on GitHub for everything and want enterprise governance baked in.

Windsurf (Codeium's editor) — A VS Code fork with a different take on AI integration than Cursor. Strong at multi-step agentic flows; integrates with Codeium's own models and others.

Cline — Open-source VS Code agent extension. Tighter integration with VS Code than a separate IDE, more agentic than Copilot. Active community.

Continue — Open-source, model-agnostic AI assistant for VS Code and JetBrains. Best for teams that want to run their own models or route to multiple providers.

Zed AI — The Zed editor's native AI features. Fast editor, polished AI integration. Smaller community than the VS Code-based tools but real momentum.

How to pick

Three questions in order:

  1. Where do you actually live? Terminal-first → Claude Code, Aider, or Codex CLI. VS Code-first → Cursor, Windsurf, or Continue. GitHub-first → Copilot + Copilot Workspace. JetBrains-first → JetBrains AI Assistant + Continue.
  2. Which models do you want to use? Cursor and Claude Code give first-class access to Claude models. Copilot routes through OpenAI and Anthropic. Aider routes to anything. If you want to use local models (Ollama, vLLM), Aider and Continue are the cleanest paths.
  3. How agentic do you want it? Inline completion only → Copilot or Continue. Multi-file edits with chat → Cursor or Cline. Full agentic flows (set a goal, let it work) → Claude Code, Cursor agent mode, Copilot Workspace, or Windsurf.

The wrong question is "which is best." The right question is "which fits my habits and my codebase."

What does not matter as much as you'd think

A few axes that get a lot of marketing attention but tend not to drive real productivity differences:

  • Specific model claims. A frontier model from a major lab is what's actually doing the work. Differences in tool quality are mostly differences in prompting and agent loop design, not raw model capability.
  • Inline-completion speed. Below a certain bar all the tools are fast enough. Beyond that, fancier completion does not generally make you faster — see ai-coding-workflows-that-actually-work.
  • The autocomplete benchmark numbers vendors quote. None of these correlate well with how productive the tool actually makes you on your real codebase.

What does matter

  • How easily it reads your repo. Tools that index the codebase well make better edits.
  • How transparent the changes are. Tools that show you the diff clearly, let you accept/reject by hunk, and integrate with git make the review loop work.
  • How well it handles your existing workflow (tests, formatters, lints, CI). The tools that respect your repo conventions ship more useful code than the ones that fight them.
  • The price under your real usage. Most tools have aggressive free tiers and steep paid tiers. Try the free tier on real work for a week before committing.

The rest of this cluster goes deeper on the most-asked-about head-to-heads and on Aider specifically because it sits in a useful spot the marketing rarely covers.