Command Line Interfaces (CLIs)

Learning Objectives

• Identify the leading AI CLI tools and their distinct positioning
• Explain why terminal-based AI development is preferred by many experienced developers
• Compare Claude Code, OpenAI Codex CLI, Gemini CLI, Aider, and Continue.dev on license, capability, and flexibility

Why CLIs Matter

Terminal-first developers often find browser-based AI tools disruptive. Opening a browser tab interrupts the flow; copying code in and out adds friction; the browser interface doesn't integrate with shell scripts, pipes, or CI.

AI CLI tools address this: they bring powerful coding assistance directly to the terminal, where experienced developers already live. The best of them go beyond assistance — they're full agentic systems that can receive a task description, read the codebase, implement changes, run tests, and push a branch, all without leaving the terminal.

For teams running automated workflows (CI, batch processing, scheduled code review), CLI tools also integrate naturally in ways browser-based tools cannot.

Claude Code — The Autonomous Coding Agent

Claude Code is Anthropic's terminal-based coding agent — the most capable autonomous coding tool available as of early 2026.

Installation and basic use:

npm install -g @anthropic-ai/claude-code
claude

Once running, Claude Code reads your local filesystem, understands your project structure, and operates as a full engineering collaborator. Unlike tools that respond to one instruction at a time, Claude Code maintains state across a session: it knows what it's tried, what worked, what failed.

Core capabilities:

Autonomous task completion: Describe a task — "implement pagination for the lessons API" — and Claude Code reads the codebase (up to 1 million tokens of context), implements the feature across however many files are needed, runs your tests, debugs failures it encounters, and presents a working result.

Voice Mode (/voice): Push-to-talk voice interaction in 20 languages — describe what you want verbally while your hands are on the keyboard.

Computer Use: On Mac, Claude Code can operate your desktop — move the mouse, click buttons, interact with browsers and applications. Useful for testing UI changes, interacting with tools that lack APIs, and end-to-end workflow verification.

/loop (Scheduled Tasks): Run prompts on a recurring interval — cron-like scheduling for PR reviews, monitoring, or any periodic task. Example: /loop 5m /review-prs checks for new PRs every 5 minutes.

Remote Control: Connect to a running Claude Code session from outside the terminal — initiate tasks and check on them remotely. Useful for headless and cloud-based workflows.

Interactive Visuals: Produce interactive charts, diagrams, visualizations, and mobile app previews directly from the terminal.

/powerup Lessons (April 2026): Interactive tutorials within the CLI that teach Claude Code features through hands-on exercises — a built-in onboarding system for new users.

Git integration: Creates branches, makes commits with descriptive messages, pushes branches, and opens pull requests — the entire git workflow without manual commands.

MCP support: Connect any MCP server to Claude Code. With Supabase MCP, Claude Code can inspect your actual database schema. With Vercel MCP, it can read deployment logs. With GitHub MCP, it can read issues, not just code.

CLAUDE.md / AGENTS.md: Place project instructions in CLAUDE.md at the root and Claude Code reads them automatically at session start.

Best for: complex, multi-file implementation tasks; debugging across system boundaries; teams that want an autonomous agent with their existing toolchain.

OpenAI Codex CLI

OpenAI's Rust-based terminal interface (v0.117.0-alpha.12 as of March 2026) brings GPT-5.5 and the full Codex model family to the command line. Install via npm i -g @openai/codex or brew install --cask codex.

The Codex CLI defaults to GPT-5.2-Codex for ChatGPT subscribers (use /model to switch to GPT-5.5). It supports:

• Device-code ChatGPT sign-in — authenticate with your existing ChatGPT account (no separate API key needed)
• Multi-turn coding conversations
• File system access and modification
• Terminal command execution with confirmation prompts

For developers who prefer working in OpenAI's ecosystem and want Codex model capability without opening a browser, this is the direct path.

Gemini CLI — Open Source with 1 million Context

Gemini CLI (v0.34.0 stable as of March 2026) is Google's open-source terminal agent, now significantly more mature:

1 million token context window: Gemini CLI can analyze an entire large codebase, an extensive documentation set, or a very long conversation history in a single context. For analyzing legacy systems or documenting large projects, this is transformative.

Plan Mode (enabled by default): Gemini CLI now plans before acting — breaking complex tasks into steps and presenting the plan for review before executing. This reduces wasted work on misunderstood tasks.

Intelligent model routing: Automatically selects between Gemini 3.1 Pro (complex reasoning) and Gemini 3 Flash (speed) based on task complexity.

gVisor sandboxing: Native sandboxing (with experimental LXC support) isolates agent actions from the host system — safer execution of untrusted operations.

Subagents (preview): Spawn sub-agents with tool isolation for parallelized work across different parts of a codebase.

Free daily usage: With a Google account, a meaningful daily usage allowance is free — relevant for individual developers and students.

npm install -g @google/gemini-cli
gemini

Best for: analyzing very large codebases, documentation tasks, cost-sensitive workflows, teams in the Google ecosystem, and developers who want Plan Mode's structured approach.

Aider — Git-Integrated, Any Model

Aider ("AI pair programmer") is an open-source CLI tool with one defining characteristic: every change the AI makes is automatically committed to git with an explanatory commit message.

This design philosophy — "AI changes should be version-controlled automatically" — has practical implications:

• Every AI action is reversible: git revert undoes it cleanly
• The git history becomes a record of AI contributions
• Reviewers can see exactly what the AI changed and why (from the commit message)

Aider works with any LLM: OpenAI, Anthropic, Google, or local models via Ollama. This model flexibility makes it attractive for teams with specific model preferences or privacy requirements.

pip install aider-install
aider --model claude-opus-4-6

Best for: teams that want AI coding assistance with strong version control discipline; developers who want model flexibility without vendor lock-in; git-heavy workflows.

Continue.dev — From IDE Extension to Continuous AI

Continue.dev has undergone a major pivot from an IDE extension to Continuous AI — an open-source CLI that runs async agents on every pull request. While the VS Code and JetBrains extensions still exist, the primary product is now a headless agent platform for CI/CD integration.

Key capabilities (post-pivot):

Headless mode (cloud): Runs in CI pipelines — agents automatically review, test, and improve every PR without developer intervention.

TUI mode (interactive): Terminal UI for direct interaction when you want to guide the agent.

Integrations: GitHub, Sentry (error monitoring), Snyk (security scanning), and CI/CD pipelines — agents can react to errors, security findings, and PR events.

Still open source and free: No paid tiers. Apache 2.0 license.

The IDE extensions remain useful for developers who want local AI coding assistance with 100+ model support (including Ollama for local models). But the strategic direction is now automated agents in the CI/CD pipeline, not interactive IDE assistance.

Best for: teams wanting automated AI on every PR; CI/CD integration with error monitoring and security scanning; developers who want a free, open-source agent platform.

Mistral Vibe CLI — Open-Source Terminal Agent

Mistral Vibe CLI (v2.0) is Mistral's open-source terminal coding agent, powered by Devstral 2.

Key capabilities:

• Custom subagents: Define specialized sub-agents via TOML configuration files — each with its own tools, context, and personality
• Multi-choice clarifications: When the task is ambiguous, Vibe presents structured options rather than guessing
• Slash-command skills: Define project-specific skills (similar to Claude Code's SKILL.md pattern)
• MCP support: Full MCP integration for connecting to external services
• Available as a Zed extension as well as a standalone CLI
• Open source, on-premise capable

Best for: Mistral ecosystem preference; teams wanting customizable sub-agents; on-premise deployments; developers using Zed as their IDE.

Goose (Block)

Goose (v1.28.0, 29,400+ GitHub stars) is an open-source agentic coding tool built by Block (the company behind Square and Cash App). In December 2025, Goose was contributed to the Agentic AI Foundation (AAIF) under the Linux Foundation — alongside MCP and AGENTS.md — backed by AWS, Anthropic, Google, Microsoft, and OpenAI.

Goose is MCP-native in its design: its capabilities come from MCP server integrations rather than being built directly into the tool. This makes it modular and extensible — add MCP servers to expand what Goose can do. Available as both a CLI and a desktop application.

Philosophy: Goose is built by developers who believe AI coding tools should be open, composable, and controlled by the developer community rather than by individual vendors. The Linux Foundation backing gives it institutional longevity and governance credibility.

Key Takeaways

• CLI tools bring agentic AI coding to the terminal, integrating with Unix workflows, git, and CI in ways browser-based tools cannot — and the field is rapidly maturing with Voice Mode, Computer Use, Plan Mode, and sandboxing
• Claude Code is the most capable autonomous coding agent (1 million context, Voice Mode, Computer Use, /loop); Gemini CLI offers Plan Mode and gVisor sandboxing; Mistral Vibe CLI adds custom subagents
• Continue.dev has pivoted from IDE extension to CI/CD-integrated async agent platform — a sign that the industry is moving toward automated agents in the development pipeline, not just interactive tools
• The open-source options (Gemini CLI, Aider, Continue.dev, Goose, Mistral Vibe CLI) provide meaningful alternatives to proprietary tools — Goose's donation to the Linux Foundation signals long-term community governance