Guide
Terminal coding agents in 2026: the CLIs worth knowing
A terminal coding agent is a CLI you point at a repository and a task: it reads, edits, runs commands and reports back. There are now enough of them that the useful skill is trying several on your own code.
From the model labs
- Claude Code — Anthropic’s agent, driven by Claude models, with a subscription or API login.
- Codex CLI — OpenAI’s open-source terminal agent.
- Gemini CLI — Google’s open-source agent for Gemini models.
- Qwen Code — an agent for Alibaba’s Qwen models, derived from Gemini CLI.
From tool and editor makers
- GitHub Copilot CLI — Copilot as a terminal agent, with a GitHub Copilot subscription.
- Cursor CLI — Cursor’s agent outside the editor.
- Amp — Sourcegraph’s coding agent.
- Droid — Factory’s agent.
Open source and model-agnostic
- Aider — one of the earliest, git-native, works with many model providers.
- opencode — an open-source terminal agent with a TUI and many providers.
- Goose — Block’s open-source, extensible agent.
- Crush — Charm’s terminal agent.
- Cline — started as an editor extension, with a CLI as well.
How to choose: run them on the same task
Benchmarks measure somebody else’s repository. The comparison that matters is your codebase and your kind of task. Give the same well-defined task to two or three agents, each in its own git worktree, and read the diffs: scope, correctness, how much unrequested change came along.
Nix Agents is built for exactly that on Windows — every CLI above runs as itself in its own worktree and terminal, side by side on one board, with your own logins. The integration pages cover each one.
Questions
- Which terminal coding agent is best?
- It depends on your codebase, your model access and the task. Running two agents on the same task in separate worktrees and comparing diffs answers it faster than any ranking.
- Can I use several coding agents on one project?
- Yes. Give each its own git worktree and branch so they cannot overwrite each other, and review each diff before merging.