[ PLAYBOOK · 03 ] · MAY 7, 2026 · 2 min
Agentic workflows with the OpenAI Codex CLI.
The OpenAI Codex CLI is the local terminal agent in OpenAI's coding family. The case for picking it over Claude Code; the case for picking neither.
The take
The Codex CLI is OpenAI's local-first agentic coding tool. It belongs in the same category as Claude Code: a CLI that reads your codebase, writes code, runs your tests, and surfaces every action for review. Pick it when you have already standardized on GPT-5.x, when source code must stay on your machine, or when you want Ollama-backed offline operation. Otherwise the choice is texture, not category.
Why
Codex CLI runs locally. Source code stays on your machine; only the prompt and high-level context go to the model. The recommended model in mid-2026 is GPT-5.5, OpenAI's frontier coding model (GPT-5.4 remains as an API-key fallback). The CLI is open source under openai/codex on GitHub, built in Rust for speed. Install with npm i -g @openai/codex or brew install --cask codex.
Two practical capabilities matter for SAES audiences. First, the local provider option: codex --oss pointed at a running Ollama instance lets you run inference on a fully local model. That removes the OpenAI dependency for sensitive workloads. Second, native Windows arrived in March 2026: PowerShell, native sandbox, no WSL required for development. OpenAI still recommends WSL2 for production-grade isolation.
When this breaks
The Codex CLI is not the Codex cloud platform. The CLI is one developer, one terminal, one task at a time. If you want a fleet of agents opening pull requests in parallel across many services, you want the cloud Codex App, not the CLI.
It also breaks for teams already invested in Anthropic's tooling stack: Claude Agent SDK, MCP servers, project-level skills. The two ecosystems are not fungible at the integration layer. Switching costs are real, especially if the team has authored its own skills.
What to do this week
If you have not committed to a coding-agent ecosystem yet, install both Codex CLI and Claude Code. Run them in parallel for one week on the same task class. Long-tail bug investigation works well. Note where each one stalls, where each one excels, and how each handles your team's review rituals.
Decide on model preference and ecosystem fit, not on benchmark videos. If your team has already standardized on Claude Code, do not switch on principle. Switch only if a specific compliance requirement (data sovereignty, model preference) forces it.