Claude Code vs Codex CLI vs Gemini CLI: The Agentic CLI Wars

April 2026: three major AI companies are competing for the agentic coding CLI. All three have converged on the same primitives. The winner isn't determined by the CLI — it's determined by the engineer who masters what's underneath all three.

Here's the practitioner breakdown.

The Three Philosophies

Claude Code (Anthropic): Terminal-Native Agentic

• Language: TypeScript. Model: Opus 4.6 / Sonnet 4.6
• Context: 1M tokens. Config: CLAUDE.md. Source: Closed (CLI open)
• Special: Computer Use built-in, Agent Teams, Scheduled Cloud Tasks
• Philosophy: Operates at the system level. Reads, writes, and executes autonomously. Skills + Hooks + MCP ecosystem. CLAUDE.md for project context.

Codex CLI (OpenAI): Speed + Open Source

• Language: Rust (speed). Model: GPT-5.4 / GPT-5.3-Codex (Fast model: GPT-5.3-Codex-Spark — 1000+ TPS)
• Context: 200K tokens. Config: AGENTS.md. Source: Open Source
• Special: Codex Cloud async tasks, @codex in GitHub PRs & Issues, config.toml
• Philosophy: VS Code fork with AI woven in. Best tab-completion experience. Composer for multi-file edits. Multi-model switching.

Gemini CLI (Google): Free + Grounded

• Language: TypeScript. Model: Gemini 3 Pro / Gemini 3.1 Pro Preview
• Context: 1M tokens. Config: GEMINI.md. Source: Open Source — Apache 2.0
• Special: Google Search grounding built-in, Browser Agent (experimental), shares quota with Code Assist
• Philosophy: Works in any IDE. Lowest friction to adopt. Deep Google ecosystem tie-in. Enterprise IP indemnity.

Feature Matrix

Pricing Reality

Gemini CLI has the most generous free tier by a wide margin. Claude Code is the most capable at the high end. Codex CLI wins on raw inference speed with its Spark model.

The Key Insight: All Three Have Converged

Despite different philosophies, all three CLIs have converged on the same primitives:

• MCP for tool integration
• Context files for project memory (CLAUDE.md / AGENTS.md / GEMINI.md)
• Subagent architectures for multi-agent workflows
• Sandboxed execution for safety
• Skills/plugin systems for extensibility

Differentiation = Model Quality × Context Depth × Ecosystem Lock-in

The universal project pattern that works for all three:

your-project/

CLAUDE.md / AGENTS.md / GEMINI.md # project context

.mcp.json # MCP server config

skills/ # reusable agent skills

hooks/ # pre/post tool execution

subagents/ # multi-agent orchestration

The Practitioner Verdict

Enterprise / Complex Codebases → Claude Code

1M context + Opus reasoning + deepest hook system. If you're refactoring 50+ files, building complex multi-agent workflows, or need maximum autonomy — this is your tool.

Speed / Open Source → Codex CLI

Rust-native + 1000+ TPS Spark + strongest cloud pipeline. If you need raw speed and want to self-host or contribute to the ecosystem.

Free / Getting Started → Gemini CLI

1K free req/day + Google Search grounding + Apache 2.0. The best entry point, especially if you're already in the Google ecosystem.

The winning combos:

• Power Stack: Cursor + Claude Code — daily coding in Cursor, heavy lifting in Claude Code
• Enterprise Stack: VS Code + Copilot + Claude Code — completions via Copilot, agentic work via Claude Code
• Budget Stack: Gemini CLI free + Claude for complex problems

The winner isn't the CLI. It's the engineer who masters the shared primitives across all three.

Jason Newell is an AI practitioner, builder, and writer covering agentic systems, developer tooling, and the future of AI engineering.