What it is
Terminal coding agents live in the command line, not in an IDE plugin. You describe a task in plain language right in the terminal, and the agent edits files, runs tests, commits and debugs without ever leaving the shell. This is the new CLI generation: Claude Code, OpenAI Codex CLI, Google Gemini CLI, OpenCode, Aider.
Where it came from
- February 2025 — Anthropic launches Claude Code as a terminal agent.
- April 16, 2025 — OpenAI ships Codex CLI alongside the o3 and o4-mini models.
- 2025 — Google rolls out Gemini CLI (generous free tier, a 1M-token window). The terminal goes from "a niche thing for hackers" to the default interface for AI coding.
Why it took off
- The terminal already sits next to the tools: git, tests, builds, the dev server — the agent needs no bridge into an IDE.
- Composability: a CLI drops into scripts, CI, pre-commit hooks and background jobs.
- Growth numbers: in roughly 4 months Claude Code reached 115,000 developers and was processing 195M lines a week (disclosed July 2025).
How to use it today
- Install one CLI agent and introduce it to your project: a
CLAUDE.md/AGENTS.mdwith test commands and conventions. - Use non-interactive mode for automation: an agent in CI fixing flaky tests or generating migrations.
- In 2026 the pros keep several around at once (Claude Code for quality, Codex for the sandbox and token economy, Gemini for the big context and free tier) and pick per task.
What to watch out for
- Shell access equals access to everything: run it inside the project directory and never hand it production credentials.
- A terminal agent easily gets carried away — set limits on steps and spend.
- Don't confuse it with editor autocomplete: this is a tool for tasks, not for typing lines.