"AI coding agent" covers tools that do very different things, and the comparison pages here needed a grouping before they could be fair. The one we settled on, with the projects that anchor each group:
Terminal pair programmers. You drive, the tool edits. Aider (Apache-2.0) is the reference: it builds a repository map so the model knows what exists without reading everything, makes edits as diffs, and commits each change to git so you can undo with git revert. It works with any model you have a key for.
Editor extensions. Cline (Apache-2.0) in VS Code, with a Plan mode (discuss) and an Act mode (edit and run), and MCP support for connecting tools. Continue was the other anchor here until its acquisition by Cursor; /compare/cline-vs-continue is still up because the code is.
Autonomous sandboxes. OpenHands (MIT, formerly OpenDevin): the agent runs in a container with its own shell and browser, and works a task to completion with less interaction. Different trust model; you review the result, not each step.
Vendor CLIs that are open source. Gemini CLI (Apache-2.0) and Codex CLI (Apache-2.0, the openai/codex repository): terminal agents from model vendors, open source but tied to their models in practice.
Pairs across the groups: /compare/aider-vs-cline, /compare/aider-vs-openhands, /compare/codex-vs-gemini-cli, /compare/aider-vs-opencode. Vocabulary: /learn/ai-coding-agent.
The grouping is ours and arguable. If a project is in the wrong box, say so here.